-2.5cmthe forced safety effect: how higher capital ...faculty.london.edu/fmalherbe/bm.pdf · how...
TRANSCRIPT
The Forced Safety Effect:How Higher Capital Requirements Can
Increase Bank Lending
Saleem Bahaj (BoE and CFM) and Frederic Malherbe (LBS and CEPR)∗
3rd July 2018
Abstract
Government guarantees generate an implicit subsidy for banks. Even though
a capital requirement reduces this subsidy, a bank may optimally respond to
a higher capital requirement by increasing lending. This requires that the
marginal loan generates positive residual cashflows in the states of the world
where the bank just defaults. Since an increase in the capital requirement
makes the bank safer, it makes the shareholders internalise such cash flows.
We dub this mechanism, which we argue is empirically relevant, the forced
safety effect.
∗Corresponding author Frederic Malherbe at: London Business School, Regent’s Park NW1 4SA, London United Kingdom,tel: +442070008428, email: [email protected]. A previous version was circulated under the title: “A Positive Analysis
of Bank Behaviour under Capital Requirements”. It drew extensively on the BoE Staff Working Paper (Bahaj et al. 2016)
that we have written with Jonathan Bridges and Cian O’Neill. We thank them for their contribution and for allowing us to
build off this previous work. We are particularly grateful to Jason Donaldson, and to Kartik Anand, Heski Bar-Isaak, Max
Bruche, Matthieu Chavaz, Filippo de Marco, Stijn Ferarri, Peter Feldhutter, Pablo Alberto Aguilar García, Tirupam Goel,
Sebastian Hohmann, David Martinez-Miera, Tom Norman, Marcus Opp, Daniel Paravisini, José-Luis Peydro, Hélène Rey,
Oleg Rubanov, Carmelo Salleo, Javier Suarez, and Emily Williams for their discussions and comments. We also thank
seminar and conference participants at the University of Zurich, the University of Bonn, the London Business School,
the Bank of England, the Bundesbank, the Bank of Greece, Banco de Portugal, the Dutch National Bank, Pompeu Fabra
University, Oxford Said Business School, the University of Amsterdam, Stockholm School of Economics, Stockholm Busi-
ness School, Toulouse School of Economics, Copenhagen Business School, Queen Mary-University of London, University
College London, Bocconi University, SciencesPo, INSEAD, EIEF, Goethe University, Frankfurt School of Finance, War-
wick Busines School, Oxford University, HEC Paris, Olin School of Business at WashU, EPFL-HEC Lausanne, Rotterdam
School of Management, Reserve Bank of New Zealand, Cass Business School, the University of Wisconsin-Madison, the
Paul Wooley Center Conference at LSE, the Baffi Carefin conference at Bocconi, 2016 ESEM Meetings, the 2018 AFA meet-
ings, the 2018 European Winter Finance Conference, and the 11th Swiss Winter Conference on Financial Intermediation
for useful feedback. The views expressed here are those of the authors and do not necessarily reflect those of the Bank of
England, the MPC, the FPC, or the PRC.
1
1 Introduction
Since the global financial crisis, bank capital requirements have been substan-
tially tightened.1 The merits of these reforms have been fiercely debated. Higher
capital requirements, the typical refrain goes, raise banks’ costs of funds, thereby
reducing credit provision and dampening economic activity.2 However, the aca-
demic literature has pointed out that bank capital is unlikely to be socially costly.
Hence, increases in banks’ private costs of funds are irrelevant from a normative
perspective (see, e.g. Hanson et al. (2011) and Admati et al. (2013)). Nonetheless,
the idea that such increases result in less lending has seeped into conventional
wisdom.
In this paper, we challenge such conventional wisdom. We develop a model
where capital is costly from a bank’s perspective due to an implicit subsidy from
a government guarantee. A higher capital requirement always reduces the value
of this subsidy. Hence, it is correct that the bank’s costs of funds go up (at a given
level of lending). However, the total value of the subsidy is not what is relevant
for the bank’s lending decision. What matters is the marginal subsidy – i.e., the
extent to which the marginal loan is subsidised. We show that an increase in the
requirement can increase the marginal subsidy. Then, facing a lower marginalcosts of funds, the bank increases lending.
Our model has a single period in which a representative bank faces a capital
requirement and finances loans with a mix of liabilities that can be interpreted as
deposits and capital. The bank starts with existing loans and can make new ones.
All loans mature at the end of the period. New loans present diminishing marginal
returns, which are not necessarily perfectly correlated with those on legacy loans.
Deposits and capital are supplied perfectly elastically by risk-neutral households.
The bank maximises the expected payoff of initial shareholders. Deposits are
insured by the government with no fee; hence, they are implicitly subsidised. This
has two implications. First, the capital requirement is binding in equilibrium: the1Specifically, minimum tier one capital requirements were raised from 4 to 6% of risk weighted
assets, but additional extra “buffers” were created to adjust, inter alia, for the systemic import-ance of the institution, the economic cycle, and to prevent accidental breaches of the minimum.Effective requirements for large global banks are now in the double digits of risk-weighted assets.
2See, for instance, Institute of International Finance (2011) on page 10.
2
bank chooses lending and adjusts capital to meet the requirement. Second, the
objective function can be written as the sum of the economic surplus from lending
and a term that captures the value of the implicit subsidy (as in Merton (1977)).
The derivative of the subsidy with respect to lending, the marginal subsidy, is
a wedge in the first order condition. This wedge captures the underlying moral
hazard problem arising from the guarantee.
We study how marginal changes in the capital requirement affects the equi-
librium level of lending (we refer to the response to an increase as the lendingresponse). By definition, such changes do not directly affect economic surplus.
Hence, if an increase in the requirement increases the marginal subsidy, the bank
increases lending as a result.
Increasing the capital requirement has two effects on the marginal subsidy.
First, a smaller fraction of the marginal loan is financed by deposits. This gen-
erates an intuitive composition effect: the bank substitutes subsidised deposits
with capital. This effect captures exactly how the capital requirement raises the
bank’s average funding costs.3
However, the change in the capital requirement also affects whether the bank
defaults or not in any given state. To go further, it is useful to define the residual
cashflow associated with the marginal loan. This variable, which we denote Z,
is what is left from the marginal loan’s realised payoff after deducting the cost of
deposits raised to finance it. In the states where the bank survives, Z comes as
an addition to the shareholder’s payoff. But if the bank defaults, Z accrues to the
taxpayer.
The second effect, which we argue is overlooked by conventional wisdom, goes
as follows. Consider the default boundary – that is, the set of states where the
bank can just repay depositors. Increasing the requirement increases the buf-
fer against losses and shifts the default boundary. There are now more states
where Z accrues to shareholders. In particular, increasing the capital require-
ment makes the shareholders internalise the expected value of Z along the de-
fault boundary. This second effect captures that the requirement forces the bank
3The change in average funding costs is relevant to determine the impact on the bank’s profit.See Kisin and Manela (2016) for a quantification.
3
towards safety.
Because the bank could have initially chosen to be safer and internalise these
cashflows (by operating at a higher capital ratio than the requirement), but pre-
ferred not to, we dub the second effect the Forced Safety Effect (FSE).
If, in expectation, the residual cashflows along the default boundary are pos-
itive, the bank is internalising cashflows that increase the shareholders’ payoff.
In this case, the FSE is positive and it increases the value of the marginal sub-
sidy. If the cashflows are negative, the FSE makes the bank internalise more
losses. Hence, it decreases the value of the marginal subsidy and reinforces the
composition effect.
Our main contribution is to show that: (i) the FSE can be positive (ii) the FSE
can dominate the composition effect, which is why lending can increase with the
capital requirement.
We solve an extended version of the model numerically to explore the empirical
relevance of our analytical results.4 We find a positive lending response in plaus-
ible conditions – namely, in a calibration that targets the situation facing a global
bank in 2017. However, at levels of capital requirements prevailing before the
global financial crisis, lending responses are more likely negative. Overall, our
sensitivity analysis reveals that lending responses are likely to exhibit substan-
tial variation both in the cross section of banks and in the time series. Hence,
one should not necessarily expect an homogeneous relationship between the cap-
ital requirement and bank lending. We argue, in Section 5.3, that this may help
reconcile results in the empirical literature, and we discuss the empirical predic-
tions that arise from our analysis in Section 6.2.
In the literature, the idea that tighter capital requirements raise banks’ average
costs of funds and prompt a credit contraction has been formalised by Thakor
(1996), among others. As Suarez (2010) discusses, a usual way to capture such
an effect is to assume an exogenous cost of issuing outside equity, in a way that
makes the capital requirement, in effect, a tax on lending. A negative lending
response then emerges naturally, as in recent quantitative studies (e.g., Corbae
4Specifically, we model competition to microfound a downward sloping demand for loans, in-troduce tax-deductible interest payments on deposits, and consider ex-ante heterogeneity amongbanks.
4
and D Erasmo (2017) and Elenev et al. (2017)).5
However, general equilibrium effects can overturn this. For instance, higher
capital requirements can reduce the aggregate supply of deposits. If deposits
provide liquidity services, this can decrease banks’ funding costs in equilibrium
and can generate a positive aggregate lending response (Begenau (2018)). We
highlight that a higher capital requirement can actually increase lending, despite
increasing average funding costs (i.e., reducing the average subsidy) in equilib-
rium.
The key ingredient for a positive FSE is a form of asset heterogeneity. The
residual cashflow of the marginal loan (i.e., Z) cannot always have the same sign
as the residual cash flow of the average loan on the bank’s balance sheet. Spe-
cifically, along the default boundary, in expectation, the percentage loss on the
marginal loan must be smaller than that on the average loan.
Introducing a second type of assets on the balance sheet (which we do with
legacy loans) makes a positive FSE possible. This second class of assets need not
be very different from the marginal loan. In fact, the returns on the two can even
be perfectly correlated, as long as percentage losses can differ. So all in all, the
heterogeneity we need is rather mild, especially with regard to the heterogeneity
of bank assets in reality.
Heterogeneity in residual cashflows generates another counterintuitive result.
To expose it and put it in perspective, it is useful to first gather insights from
different strands of literature.
Limited liability introduces a kink into the shareholders’ payoff function. Along-
side a friction, the resulting convexity can induce risk loving behaviour. For
instance, a firm that has existing debt (with given interest payments) has an in-
centive to make excessively risky investments (Jensen and Meckling (1976)), and
may value assets over their fundamental values (Allen and Gale (2000)). Such
phenomena are usually referred to as risk-shifting problems. On the other hand,
the presence of existing debt can also lead to under-investment (Myers (1977)) –
a phenomenon referred to as a debt overhang problem.
5See also Estrella (2004), Repullo and Suarez (2013), Martinez-Miera and Suarez (2014), andMalherbe (2015) for analyses of the underlying mechanisms.
5
Government guarantees to banks also generate risk-shifting behaviour: Kareken
and Wallace (1978) establish the counterpart of Jensen and Meckling (1976).
Also, when assets are loans, overvaluation leads to overlending (McKinnon and
Pill (1997)), which is the counterpart of Allen and Gale (2000).6
Our second contribution is to point out that the implicit subsidy from govern-
ment guarantees can also lead to underlending. How can a subsidy lead to less
lending? The answer is that while the total subsidy is always positive, the mar-
ginal subsidy can be negative and, therefore, act like a tax. This happens when
the marginal new loan generates positive residual cashflows in default states.
These cashflows then accrue to the taxpayer and, hence, reduce the value of the
implicit subsidy.7 We argue that this is the counterpart of Myers (1977).8 As
with the debt overhang problem, the bank undervalues new loans because of an
implicit transfer from the shareholders to other stakeholders. However, the mech-
anism is different. It does not hinge on existing debt: it actually operates if the
bank is initially 100% equity financed.9 We dub this phenomenon the guaranteeoverhang problem.10
Finally, a positive lending response and the guarantee overhang problem are
linked: they both arise from the same moral hazard problem (i.e., the guarantee)
and both require heterogeneity in residual cashflows. But we wish to stress that
they are distinct phenomena: a positive lending response can arise when the bank
overlends and a negative lending response can arise when the bank underlends.
To sum up, the punchline of this paper is that the lending response can be
positive. This happens when the FSE more than offsets the composition effect,
6See also, e.g., Rochet (1992), Martinez-Miera and Suarez 2014 on bank risk-shifting, andKrugman (1999) and Malherbe (2015) on over-investment, which is ultimately the consequenceof overlending.
7In a sense, positive residual cashflows in default provide insurance. But shareholders do notvalue insurance, as the benefits go to the taxpayer. Hence, when the bank refrains from issuinga positive NPV loan, this must be interpreted here as a way to take excessive risk.
8Papers that link bank underlending to the debt overhang problem include Hanson et al.(2011), Admati et al. (2018), and Jakucionyte and van Wijnbergen (2018).
9See Section 6.4 for more details.10In Harris et al. (2017) and Martinez-Miera and Suarez (2014), banks face a menu of risks
and could, in principle, hold assets with the relevant residual cashflow heterogeneity. However,in equilibrium, they choose not to. Underlending can happen in their models, but this is due tothe scarcity of bank capital in the aggregate, not to the guarantee overhang effect.
6
which we argue can happen in plausible conditions. However, the main takeaway
for the policy debate goes beyond the sign of the lending response. Indeed, in our
calibrations, when the FSE does not dominate, it still makes the lending response
substantially less negative than otherwise. Overlooking this effect is tantamount
to confusing how average, rather than marginal, costs of funds are affected by
changes in capital requirements.
2 The baseline model
2.1 The environment
There are two dates, 1 and 2. There is a bank and a continuum of households
who own the bank’s liabilities, and a government. Households are risk neutral
and do not discount the future; they supply funds perfectly elastically with an
opportunity cost of funding of 1. We focus on the date-1 decision of the bank.
The random variable A captures the realised state of the economy at date-2. It
is distributed according to a function f(A) with positive support [aL, aH ] and with
E[A] = 1. Figure 1 summaries the bank’s balance sheet.
Predetermined variables. As of date-1, there are legacy loans on the bank’s
balance sheet. Their total book value is λ, and they generate a risky date-2 payoff
Aλ. Without loss of generality, the bank holds no cash. The bank has existing
deposits that can be withdrawn at par at date-1 and, hence, for the analysis it is
only necessary to consider the end date-1 level deposits. The book value of capital
at the beginning of date-1 is denoted by κ.
Decision variables. The bank decides how much to lend. We denote the total
amount of new lending by x ≥ 0. For simplicity, new loans also mature at date-2.
The bank has some market power over borrowers. We capture this with a payoff
function X(x), which is increasing and strictly concave in x, with X(0) = 0. We
assume X is twice differentiable in the strictly positive domain and lim x→0Xx(x) =
∞. For now, the returns to new loans are deterministic. This helps isolate the
7
main mechanism and provide intuition. We introduce stochastic new loan returns
in Section 4.1.
At the same time, the bank adjusts its liabilities: capital and deposits. We
denote the change in capital c. The change in capital can be negative (as long
as κ + c > 0). In this case, the change can be interpreted as a dividend payment
or the value of a share repurchase. If c is positive, it should be interpreted as
the bank raising more capital. In this case, an amount c is raised in exchange
for date-2 cashflow rights. The corresponding total repayment is denoted C. This
repayment is determined in equilibrium and can be contingent on any realised
variable.
The bank’s chosen level of deposits is denoted d. As we will see, the presence
of an implicit subsidy makes deposits the most attractive source of funds for
the bank. Hence, the capital requirement will be binding in equilibrium and the
optimal choice for x will, effectively, pin down the liability side of the balance
sheet.
Deposit insurance and the capital requirement. The government insures bank
deposits with no premium: in the event the bank has insufficient cashflows to re-
pay depositors in period 2, the government makes depositors whole, and breaks
even via an ex-post lump-sum tax on households. This is the source of moral
hazard in the model. Deposits pay no interest. If the bank defaults on deposits,
no payment to any other liability is allowed.
The bank faces a capital requirement constraint that takes the form:
κ+ c ≥ γ(x+ λ), (1)
where γ ∈ (0, 1) is a parameter (which we refer to as the requirement) set by the
government, x+ λ is the book value of total assets on the balance sheet, and κ+ c
is the bank’s total capital at book value.
To be allowed to operate at date-1, the bank must satisfy the capital require-
ment; otherwise, the government shuts down the bank. In this case, initial share-
holders walk away with 0 and we impose x = c = 0.
8
Figure 1: The bank’s balance sheet
Assets Liabilities(new loans) x κ+ c = γ(x+ λ) (capital)
(legacy loans) λ d = (1− γ)(x+ λ) (deposits)Notes: The parameter γ denotes the capital requirement, κ is existing capital and c is net issuance (can be negative). Thecapital requirement is always binding; see Section 2.2.
Finally, note that the assumption E[A] = 1 ensures that initial shareholders
always find it profitable not to walk away at date-1. We explore date-1 bank
closure in Section A.1.
2.2 Setting up the analysis
Date-2 default on deposits. If date-2 cashflows are too low to repay the depos-
itors, the bank defaults on them. This happens when
d > X + Aλ
promised repayment total cash flow
We can then define a0 as the realisation for A below which the bank defaults on
deposits:
a0 ≡d−Xλ
.
We refer to a0 as the default boundary. We can also define p, the probability that
the bank does not default:
p ≡∫ aH
a0
f(A)dA.
Pricing of new capital. Investors act competitively, so that, in equilibrium, they
just break even in expectation. First, we assume that the bank issues new capital
(i.e., c ≥ 0). Denoting C(A) the contingent, date-2 repayment to new capital, we
then have: ∫ aH
aL
C(A)f(A)dA = c. (2)
9
To be able to interpret c as capital issuance, the underlying securities should be
junior to deposits. Hence, we impose
C(A) ≤ 0, ∀A ≤ a0. (3)
We also impose limited liability for investors, which implies C(A) ≥ 0, and for
initial shareholders, which implies C(A) ≤ X +Aλ− d. However, we do not restrict
new capital to be a particular form of security. What matters is that capital is
junior to deposits and will, therefore, absorb losses. In practice, one can, for
instance, think of it as seasoned equity or subordinated debt.11
Finally note that, even though they all refer to households, we use different
terms for holders of different bank-issued liabilities. Initial shareholders own
the initial (i.e., inside) equity, investors hold new capital, and depositors hold
deposits.
Initial shareholders’ payoff. If c is positive, the expected payoff to (or expected
final wealth of) initial shareholders is:
w ≡∫ aH
a0
[X(x) + Aλ− d− C(A)] f(A)dA
substituting break even condition (2) gives:
w ≡∫ aH
a0
[X(x) + Aλ− d] f(A)dA− c
Now, if c is negative, the payoff is identical to the above as initial shareholders
will receive −c with certainty in period 1. In the absence of frictions affecting the
contracting between initial shareholders and investors in new capital, the shadow
11For subordinated debt, the interest payment should compensate the loss of capital that hap-pens in some states. Assume the bank can fully repay deposits with probability p, then theexpected return for the subordinated debt holders in these states should be 1
p (and zero other-wise). In the case of seasoned equity, the logic is the same, but the mapping goes as follows: atdate-1 the bank starts with κ shares and issues κ′ additional shares at a unit price v in exchangeof c = κ′v. This gives investors the right to a payoff of C = κ′
κ+κ′ max{0, V } where V = X + Aλ − d.Hence, their break-even condition is v = E
[max{0,V }κ+κ′
].
10
value of initial capital is equal to the price of new capital. Hence, it is unnecessary
to treat positive and negative c as separate cases in what follows.12 And, accord-
ingly, there is no need to distinguish between the owners of different classes of
bank capital. For simplicity, we refer to them (both the initial shareholders and
the investors in new capital) collectively as “the shareholders”.
The problem of the bank If the bank is fully safe in equilibrium (i.e., p = 1),
the capital ratio is irrelevant (in a Modigliani and Miller sense). In this case, the
bank is locally indifferent between any mix of capital and deposits that satisfies
the requirement. For most of the analysis, we focus on the cases where the
bank defaults at date-2 with strictly positive probability in equilibrium. In these
cases, the capital requirement always binds because, from the bank’s point of
view, deposits are cheaper (depositors always break even, but sometimes at the
expense of the taxpayer). Hence, the bank’s problem boils down to finding a level
of lending x∗ that solves
maxx≥0
∫ aH
a0
[X(x) + Aλ− (1− γ) (x+ λ)] f(A)dA− (γ (x+ λ)− κ) . (4)
We refer to x∗ as the equilibrium level of lending.
3 Analysis of the baseline model
Our main result is that x∗ may increase with γ. In the case where the problem of
the bank (4) admits a unique maximum, x∗ can be expressed as a function of γ.
Then, we formalise this result as follows:
Proposition 1. For all γ ∈ (0, 1) and an associated function x∗(γ), if p (x∗(γ), γ) < 1,there exists γ′ > γ such that x∗(γ′) > x∗(γ).
12In Appendix A.1, we relax the assumption that E[A] = 1. In that case, the initial shareholder’sparticipation constraint might bind. As a result, initial and new capital are not equivalent. Oneimplication of this is that when E[A] is sufficiently low, an increase in capital requirement mayresult in the shareholders closing the bank instead of raising capital.
11
Proof. All proofs are in Appendix A. (Generalising this result to the case of multiple
maxima only complicates notation.)
To study situations where the requirement is relevant, we assume p∗(x∗(γ), γ) <
1 for what follows. For better readability, we often omit function dependencies
on x and γ. Additionally, we use subscripts for partial derivatives in these two
variables and stars to indicate where functions are evaluated in equilibrium. For
instance: p∗x ≡ px (x∗, γ).
3.1 The first order approach
We derive intuition by focusing on small increases in γ and assume that the first
order condition uniquely pins down a function x∗(γ).
The implicit subsidy decomposition
Lemma 1. The bank’s objective function can be rewritten:
w(x) = X − x︸ ︷︷ ︸economic surplus
+
∫ a0
aL
((1− γ) (x+ λ)−X − Aλ) f(A)dA︸ ︷︷ ︸≡s(x,γ), i.e. the implicit subsidy
+κ (5)
The first term in equation (5) captures, intuitively, the economic surplus gen-
erated by new loans.13 The second term integrates, over all the default states,
the difference between the promised repayment to the depositors, (1− γ) (x+ λ),
and the total cashflow available to the bank, X + Aλ. Under unlimited liabil-
ity, this term would be the expectation of how much, ex-post, the shareholders
would have to pay into the bank to make depositors whole. But instead, here,
the taxpayer is footing the bill. This is why s should be interpreted as the implicit
subsidy to the bank’s shareholders arising from the government guarantee. The
implicit subsidy is the source of a moral hazard in the model.
Remark 1. The implicit subsidy corresponds to the expected net worth of the
bank, when it is negative. As Merton (1977) has shown, deposit insurance can be13Since we assume E[A] = 1, legacy loans are valued on the balance sheet at their expected
value. Hence, (E[A]− 1)λ = 0, and they do not appear in (5).
12
interpreted as a (free, or at least mispriced) put option on the bank equity with a
strike price of 0. The implicit subsidy is therefore equal to the value of such an
option.14
The sign of the lending response The first order condition can be written as:15
(X∗x − 1)︸ ︷︷ ︸surplus maximisation
+ (1− p∗) [(1− γ)−X∗x]︸ ︷︷ ︸≡s∗x, i.e. the marginal subsidy
= 0. (6)
The first term represents economic surplus maximisation. Absent the implicit
subsidy, the bank would choose a level of lending xMM such that Xx(xMM) = 1.
This would correspond to Proposition 3 in Modigliani and Miller (1958): equilib-
rium investment (here lending) only depends on the opportunity cost of funds in
the economy. We therefore sometimes refer to xMM as the Modigliani and Miller
(or MM) level of lending.
The function x∗(γ) is generally not explicit. We can, however, establish three
of its properties. The first is that when γ is sufficiently large, x∗ = xMM . To see
this, note that when γ is large enough, a∗0 ≤ aL. Hence, p∗ = 1, and the distortion
disappears (s∗ = s∗x = 0). The second property is that, for lower values of γ such
that the bank defaults with positive probability in equilibrium, x∗ < xMM . To see
this, rewrite the first order condition as
X∗x = (1− γ) +γ
p∗,
and note that p∗ < 1 implies X∗x > 1. These two properties are in fact sufficient
to establish that there will be values of γ where x∗(γ) is increasing. The third
property is that when γ = 0, we also have x∗ = xMM . The intuition is that the
marginal subsidy is, in this special case, proportional to the marginal economic
surplus, and hence it does not distort the bank’s decision.16
14If there were not government guarantees, depositors would require an interest payment thatwould, in equilibrium, just compensate for the value of s. Hence, the objective function would boildown to X − x.
15See the proof of Proposition 1.16The first order condition shrinks here to p (Xx − 1) = 0. This result is specific to having
13
Figure 2: Example of x∗(γ) in the baseline model
Notes: The solid red line provides an example of x∗(γ), which is obtained when X is isoelastic and A is uniformlydistributed. The dotted line is the MM level of lending. Based on a numerical analysis, we conjecture that well behavedU-shapes are also obtained for general X functions and single peaked distributions, so long as the median of A is greaterthan 1− γ.
An example for how these three properties play out is given in Figure 2, where
x∗(γ) is a well behaved U shape. This enables the visualisation of Proposition
1. Either we are in an upward sloping region, and the result applies to small
increases in γ. Or, we are in a downward sloping region, and a sufficiently large
increase in γ is needed to increase lending.
Figure 2 is also useful to understand our next proposition. Since economic
surplus maximisation is not affected by γ, the gap between xMM and x∗ is solely
due to s∗x, which is the wedge in the first order condition. This wedge is negative,
which is something we will return to in Section 3.3. For now, what is important
is not the sign of the wedge, but whether it increases with γ. Intuitively, if an
increase in γ increases the extent to which the marginal loan is subsidised, the
lending response is positive (that is: dx∗
dγ> 0). Formally:
deterministic returns on new loans. As we will see in Section 4.1, when new loans bear risk, x∗(0)is generally different from xMM .
14
Proposition 2. (The sign of the lending response)
dx∗
dγS 0⇔ s∗xγ S 0
Bank capital ratio and profitability At this point, it is useful to point out that
an increase in capital requirement unambiguously decreases shareholders’ ex-
pected payoff: ∀x, wγ < 0. This is because, for any level of lending x, total funding
costs increase with γ (intuitively, the expected transfer from the taxpayer shrinks
as the share of deposits in the bank’s liabilities goes down). Formally:
wγ = sγ = −(1− p)(x+ λ). (7)
But this is not incompatible with the marginal subsidy being increasing in γ.
Figure 3 proposes a simple illustration. The solid line depicts the payoff function
associated with an initial capital requirement γ. That wγ < 0 means that the
payoff function associated with a higher capital requirement (denoted γ′) is below
the initial one. But it does not tell us whether it peaks to the left or to the right
of the initial optimum. If s∗xγ > 0, it peaks to its right. This is what happens when
x∗(γ) is upward sloping.
Finally, γ only sets a minimum. The bank is, therefore, allowed to operate
at any capital ratio γ′> γ. If the bank did so, it would then choose the level of
lending x∗(γ′). Such an option is, however, not optimal from the shareholders’
perspective. They are always better off maximising leverage (and, in the example
of Figure 3, choosing a lower level of lending than x∗(γ′)).
3.2 The forced safety effect
We have established that the sign of the lending response is that of s∗xγ, and we
have shown that it is positive for some region of γ. Now, we turn to the underlying
economic mechanism.
Property rights and the marginal residual cash flow Issuing the marginal
loan affects the bank’s cashflows. Irrespective of A, the bank’s revenue increases
15
Figure 3: The effect of an increase in γ on the bank’s objective
Notes: This diagram shows how the bank’s objective, w(x), is shifted by an increase in the capital requirement when thelending response is positive at the initial equilibrium, x∗(γ). If, instead, the lending response is negative, w(x, γ
′) peaks
to the left of x∗(γ).
by Xx, and the repayment due to depositors increase by 1 − γ. Let us define Z ,
the residual (i.e., net of deposits) cash flow associated with the marginal loan, as
the difference between the two:
Z(x) ≡ Xx − (1− γ) .
Since X∗x > 1 , the residual cash flow is positive in equilibrium: Z∗ > 0.
Now, which stakeholder is entitled to Z∗ depends on the realisation of A. If
the bank survives, the shareholders are the residual claimants. But if the bank
defaults, shareholders walk away with zero, and the taxpayer becomes, in effect,
the residual claimant.17 What determines the bank’s survival is the sign of the
total residual cash flow (Aλ+X − (1− γ)(x+ λ)), which is different from Z∗.
As we will see, Z∗ plays a key role in our analysis. In the baseline model, it is
deterministic, and always positive. This helps explain how the main mechanism
works and will also be useful to gain intuition in the full version of the model in
17Technically, the taxpayer is a claimant in the sense that there is reduction in the transferneeded to make depositors whole.
16
Section 4.
The cross-partial derivative of the subsidy The marginal subsidy is s∗x = (1 −p∗) [(1− γ)−X∗x]. Deriving with respect to γ and using the definition above, we
obtain:
s∗xγ = −(1− p∗)︸ ︷︷ ︸composition effect
+ p∗γZ∗︸ ︷︷ ︸
Forced Safety Effect
. (8)
The cross-partial derivative is the sum of two components. The first term is
negative. Raising γ reduces the portion of the marginal loan that is financed
with (subsidised) deposits. Since the bank must substitute deposits (which it
repays with probability p∗) with capital (which it always repays in expectation),
this change in the composition of liabilities reduces the marginal subsidy. We
dub this effect the composition effect.The second component captures that, keeping x∗ constant, an increase in γ
makes the bank safer: p∗γ > 0. This corresponds to a shift in the default bound-
ary, a∗0, which means that there are states of the world where the bank would have
defaulted if not for the extra capital. In these states, the rights to the residual
cashflow from the marginal loan (Z∗) switch from the taxpayer to the sharehold-
ers. In expectation, this raises the marginal subsidy by p∗γZ∗. Since this term
stems from the fact that the bank is forced to be safer (something it could have
always chosen to do), we dub this effect the forced safety effect. To the best of our
knowledge, this paper is the first to highlight such a mechanism.
Remark 2. The forced safety effect is completely different from the mechanism
behind the usual risk premium argument based upon the second proposition
in Modigliani and Miller (1958). In short, that argument is: a higher capital
ratio decreases the volatility of the return to capital (equity, for instance) and,
therefore, its required return. In our model, all agents are risk neutral, and the
equivalent notion of required return to capital is always 1, irrespective of γ.18
18Note that our analysis could equally be done in terms of a risk-neutral probability measure.So, introducing risk aversion would not affect the logic behind the forced safety effect.
17
Average versus marginal subsidy and conventional wisdom The composition
effect is behind the claim (made repeatedly by bank lobbies and sometimes by
policymakers (see for instance Brooke et al. (2015), p5)) that increasing capital
requirements would: (i) increase bank funding costs; and (ii) naturally lead to
less lending.
In the context of our model, the first part of this claim is correct. As we have
established above – the total subsidy decreases with the capital requirement, that
is sγ < 0. Of course this also applies to the average subsidy:
∂
∂γ
(s
x+ λ
)=
sγx+ λ
= −(1− p). (9)
Note that this effect corresponds exactly to the composition effect.
However, the second part of the claim, regarding lending, is incorrect. Our
interpretation is that it confounds the effect of the capital requirement on the
average and the marginal subsidy.19 Comparing equations (8) and (9) makes
clear that making this mistake is equivalent to ignoring the forced safety effect.
Positive lending response Equations (8) also makes clear that, for the lending
response to be positive, the FSE needs to more than offset the composition effect.
Proposition 1 establishes that, in our baseline model, there always is a range of
γ where it is case. The general properties of x∗(γ) make this clear: the bank gen-
erally underlends (x∗ < xMM ), but chooses x∗ = xMM for sufficiently large values of
γ. How is it possible then that an implicit subsidy distorts the outcome towards
less lending? This is the question we explore next.
3.3 A subsidy or a tax?
To reconcile the idea of an overall positive subsidy, occurring alongside a distor-
tion toward less lending, one can think of the subsidy as the sum of two compon-
19Imagine if, instead of guaranteeing deposits, the government explicitly subsidised them at aflat rate of δ. In this case, the total subsidy would be (1 − γ)(x + λ)δ. Since this subsidy is linearin x, average and marginal subsidy coincide, and are affected in the same, negative way by anincrease in γ.
18
ents – a lump-sum subsidy and a tax on lending.
The bank has legacy loans that are funded with deposits. These legacy loans
are risky and, in certain states of the world, generate negative residual cashflows
that need to be covered by the taxpayer. Specifically, the expected transfer from
the taxpayer (i.e., the total subsidy) if the bank does not issue any new loans (i.e.,
x = 0 hence a0 = (1− γ)) is:
s(0) ≡∫ (1−γ)
aL
((1− γ)λ− Aλ) f(A)dA > 0. (10)
As established above, issuing new loans generates positive residual cashflows.
Formally: ∀x ≤ x∗, Z(x) > 0 (since it is true for the marginal loan, it is also
true for infra-marginal ones). When the bank goes bust, these residual cashflows
accrue to the taxpayer. This reduces the transfer needed to make the deposits
that financed the legacy assets whole. Hence, new loans reduce the size of the
total subsidy; the marginal subsidy is negative. Formally, ∀x ≤ x∗, sx(x) = (1 −p(x))Z(x) < 0.
Since a negative marginal subsidy can be interpreted as a tax, we can write
the total subsidy as the sum of a lump-sum subsidy and a tax that increases with
lending:
s∗ = s(0) +
∫ x∗
0
sx(x)︸ ︷︷ ︸<0
dx.
The total value of the subsidy is always positive, but it is maximised at x =
0. Hence, the bank faces the following trade off: while issuing loans increases
economic surplus, it also decreases the total value of the subsidy. This is why
the bank’s lending is below xMM .
A guarantee overhang: similarities and differences with the debt overhangproblem To gain further intuition it is also useful to draw some parallels to the
debt overhang problem (Myers (1977)).
In Myers (1977), a firm has existing risky assets and existing debt on which
it will default in some states of the world. The firm considers raising capital to
19
finance a positive net-present-value project. The problem is that the cashflow
from the project will go to the existing debtholders in the default states. Hence,
there is an implicit transfer, and the firm does not capture the full expected
surplus from the project. As a result, the firm may find it preferable to pass on
this investment opportunity.
Our baseline model shares this outcome: in equilibrium, the bank passes on
positive net-present-value loans. Likewise, an implicit transfer is involved. When
the bank goes bust, the residual cashflows from the marginal loan decreases the
transfer needed from the taxpayer to make depositors whole. Hence, in expecta-
tion, part of the residual cashflows generated by the marginal loan is transferred
to the taxpayer (which resonates with our tax interpretation above), and is not
captured by the bank shareholders. This is an overhang problem, in the sense
that the existing balance sheet affects the decision to make an otherwise inde-
pendent investment.
However, it is not a debt overhang problem. The simplest way to illustrate this
is to consider a bank that is, initially, fully funded with capital (i.e., κ = λ). Noth-
ing would change in the main analysis; in particular, x∗(γ) would be unchanged.
Rather than being due to the presence of existing debt, the inefficiency is due to
the implicit subsidy that arises from government guarantees, which applies to all
deposits, existing and new. In this sense, there is a guarantee overhang.
To our knowledge, this paper is the first to make the point that the implicit
subsidy from government guarantees can act as a tax at the margin and, as a
result, generate such an overhang problem.
4 Extensions and the full model
In this section, we extend the baseline model to derive further insight into the
determinants of the lending response. We then combine all these ingredients in a
full model that we calibrate in Section 5.
20
Figure 4: Two examples of x∗(γ) when new loans are risky
Notes: This Figure provides two examples of x∗(γ) when new loans are also risky (blue solid and blue dashed lined). Seethe main text for a description of how these two examples come about. The U-shaped red solid line is an example of x∗(γ)in the baseline model for comparison. The dotted line is the MM level of lending.
4.1 Risky new loans
4.1.1 Preamble
When new loans are risky, x∗(γ) can take different shapes. To set the stage for
this section, let us illustrate this idea with two new examples (see Figure 4).
Example 1, depicted by the solid blue line, captures the pattern that we typ-
ically find in our calibrations (See Section 5). It resembles the U-shape of the
baseline mode, but it exhibits overlending (x∗ > xMM ) and a strongly negative
lending response at low levels of γ. Example 2, depicted by the dashed blue line,
captures a somewhat more extreme case: the bank overlends, and the lending
response is negative for all γ. We obtain such a shape if, for instance, legacy
loans are safe, and A = 1.
In what follows, we show that it is the structure of residual cashflows drives
the shape of x∗(γ). In particular, what is needed for the existence of a positive
21
lending response is a form of residual cashflow heterogeneity. To establish this
concretely, we revisit the concepts of FSE and guarantee overhang to (i) formally
show that a positive lending response and/or underlending can still occur when
new loans are risky ; (ii) identify the conditions under which these will or will
not arise; and (iii) stress that these are disjoint phenomena: a positive lending
response can arise when the bank overlends, and vice versa.
4.1.2 The first order approach with a second source of risk
To make new loans risky, we introduce a new random variable B, with positive
support [bL, bH ] and E[B] = 1. The payoff function for new loans is now BX(x). Let
the joint density between A and B be given by f(A,B). We can then define two
functions, either of which can be used to define the default boundary:
a0(B) =(1− γ)(x+ λ)−BX
λ, (11)
b0(A) =(1− γ)(x+ λ)− Aλ
X, (12)
and the probability that the bank does not default as p(x, γ) =∫ b0(aL)
bL
∫ a0(B)
aLf(A,B)dAdB.
Note that E[B] = 1 preserves the definition of economic surplus, and the impli-
cit subsidy now reads:
s(x, γ) =
∫ b0(aL)
bL
∫ a0(B)
aL
((1− γ) (x+ λ)−BX (x)− Aλ) f(A,B)dAdB.
We first emphasise that the bank’s first order condition remains X∗x − 1 + s∗x = 0,
and Proposition 2 still holds: dx∗
dγ> 0 ⇔ s∗xγ > 0. However, the expression s∗xγ now
reflects the fact that marginal residual cashflows are stochastic.
22
Figure 5: Default region and boundary
(a) (b)
Notes: This Figure illustrates the default boundary and default region in the state space {[aL, aH ]× [aL, bH ]}. The defaultboundary is the locus of points such that BX∗ +Aλ− (1− γ) (x∗ + λ) = 0; the default region corresponds to BX∗ +Aλ−(1− γ) (x∗ + λ) < 0. The term Z∗ = BX∗x− (1−γ) is the equilibrium residual cashflow on the marginal loan. The thresholdb̂ = 1−γ
X∗xis such that B > b̂⇒ Z∗ > 0.
4.1.3 The FSE revisited
The cross partial derivative now reads:
s∗xγ = −(1− p∗)︸ ︷︷ ︸<0
+ p∗γz∗∆0︸ ︷︷ ︸
FSE≷0
(13)
where
z∗∆0≡ E [Z∗ | A = a∗0(B)] =
∫ b∗0(aL)
bLZ∗f(a∗0(B), B)dB∫ b∗0(aL)
bLf(a∗0(B), B)dB
is the expected marginal residual cash flow conditional on being on the default
boundary.
Panel (a) in Figure 5 depicts the default boundary (the thick black line) and the
whole default region (the red triangle) as subsets of the state space {[aL, aH ]× [bL, bH ]}.The subscript ∆0 in z∗∆0
, refers to the fact that the default boundary is what pins
down the triangular default region.
As in the baseline model, raising γ shifts the default boundary. Hence, at each
23
point on the initial boundary, Z∗ now accrues to the shareholders. In expectation,
this additional residual cash flow is worth p∗γz∗∆0
. As before, p∗γ > 0. Whether the
FSE is positive or not therefore hinges on the sign of z∗∆0. We have:
Proposition 3. (i) z∗∆0can be positive; (ii) this implies a positive forced safety effect;
(iii) and can lead to a positive lending response: s∗xγ > 0.
We mentioned above that a positive FSE requires some heterogeneity in resid-
ual cashflows. We can now be precise about what this means. By construction,
on the default boundary, the residual cashflow on the bank’s average loan is nil
(BX∗ + Aλ − (1− γ) (x∗ + λ) = 0). If all loans where homogeneous, we would also
have Z∗ = 0 at all points on the boundary, and the FSE would be nil.
How should one interpret z∗∆0> 0? The condition means that that, in expecta-
tion along the default boundary, the marginal loan fares better than the average
loan (which, by definition, have zero residual cashflows on the boundary). In
Panel (b) of Figure 5, Z∗ > 0 in all the states above the threshold b̂. Hence, z∗∆0will
be positive if there is sufficient probability mass concentrated on the correspond-
ing upper segment of the boundary.
Since X is concave, the marginal loan is the worst, among new loans. Hence,
in our model, the marginal loan can only fare better than the average loan (i.e. av-
erage over the the whole balance sheet) if legacy loans generate negative residual
cashflows in those states. This is why legacy loans play an important role in our
model: they provide the necessary heterogeneity in residual cashflows.20 That
said, as we discuss in Section 6.1, there is nothing special about legacy loans:
there are many potential other sources of residual cashflow heterogeneity. In
general, the key necessary condition for z∗∆0> 0 to be possible is that the marginal
loan sometimes does not perform too badly when the bank defaults.
How do the model parameters affect the FSE, and ultimately, the lending re-
sponse? Both the default boundary and the marginal loan are endogenous ob-
jects. Tractability is an issue for further comparative statics, but we can still
20If A = 1, as in Example 1 above, legacy loans always generate strictly positive residual cash-flows. The marginal loan always fares worse, in expectation on the default boundary, than theaverage loan and a positive lending response is therefore not possible.
24
make a couple of points that will be useful for understanding the results of our
calibration exercises below.
First, increasing the downside of legacy loans means that they will generate
more negative residual cashflows. Ceteris paribus, decreasing the mean of A or
increasing its variance is likely to increase the FSE.
Second, lower dependence between A and B makes it less likely that Z∗ is
negative when legacy loans perform badly. Hence, this can contribute to a large
FSE too.
Before turning to the calibration exercise, let us also revisit the guarantee
overhang and its subtle relation to the FSE.
4.1.4 The guarantee overhang revisited
The expression for the marginal subsidy now reads:
s∗x = − (1− p∗) z∗∆︸︷︷︸≷0
,
where
z∗∆ ≡ E [Z∗ | A ≤ a∗0(B)] =
∫ b∗0(aL)
bL
∫ a∗0(B)
aLZ∗f(A,B)dAdB
1− p∗
is the expected marginal residual cash flow over the entire default region (see
Figure 5). As we note, z∗∆ and, therefore, s∗x can now have either sign.
If z∗∆ is negative the marginal loan will increase, on average, the required trans-
fer from the taxpayer to make depositors whole when the bank defaults. Hence,
the marginal loan benefits from a positive subsidy, and the bank overlends relat-
ive to xMM . Here, the marginal subsidy really acts as a subsidy.
In panel (b) in Figure 5, z∗∆ > 0 if there is sufficient probability mass concen-
trated in the area of the default region that is above the b̂ threshold. In that case,
the subsidy again acts as a tax, and the equilibrium is symptomatic of the guar-
antee overhang problem that we identified in the baseline model. We therefore
have a result that shares the logic of Proposition 1:
Proposition 4. For all γ such that s∗x(γ) < 0, there exists γ′ ≥ γ such that x∗(γ′) >x∗(γ).
25
Proposition 4 only expresses a sufficient condition. Perhaps counter-intuitively,
equilibrium overlending – that is, s∗x > 0 – is compatible with a positive FSE. One
can even construct cases where the FSE dominates and the lending response is
positive when the bank overlends. Figure 10 in Appendix C provides a numer-
ical example. To generate this, we used a joint distribution, f(A,B), with a lot
of probability mass concentrated in the lower left corner of the default region –
i.e., A and B have high tail dependency. But we restricted A and B to have a low
correlation away from the lower tail, so the dependency between A and B is weak
along the default boundary.
While such cases are a bit extreme, and less likely to be relevant empirically,
they reinforce the point that z∗∆ and z∗∆0are different objects and can have different
signs. Whether the bank under- or overlends does not dictate the sign of the
lending response.
4.2 Taxes and tax shields
Another reason why banks may find capital relatively costly is the tax advantage
of debt. To capture this, we assume that the bank faces a tax rate τ on positive
profits, net of interest expenses on deposits. In order to introduce a meaningful
tax shield, we also assume that households have an opportunity cost of funds
1+ρ > 1, so that the interest rate on deposits is ρ. To maintain our normalisation,
we now set E[A] = 1 + ρ.
We describe how the tax interacts with the model formally in Appendix A.2.
However, as an illustration, Figure 6 shows the two main ways the tax affects
our results. The solid red line depicts x∗(γ) without the tax (τ = 0). This is the
U-shape relationship of the baseline model. The blue dashed curve is the case
with the tax. The net effect of taxes on the lending response is ambiguous: the
U-shape relationship is still present, but it is (i) tilted clockwise and (ii) deeper
than it would be in the absence of a tax. The tilt is intuitive, it comes from
the tax deductibility: similar to the composition effect above, an increase in γ
increases the average cost of funds. The deeper U-shape is more subtle. This is
a result of the fact that the tax itself tends to reduce equilibrium lending and,
26
Figure 6: The shape of x∗(γ) with corporate income tax
Notes: The diagram provides an example of x∗(γ) in the baseline model (red solid line) and when the bank faces a corporateincome tax τ (blue dotted line). As in Figure 2, this is obtained when X is isoelastic and A is uniform. Similar patternsarise with alternative functional forms and distributions. The horizontal dotted line is the MM level of lending. Thedownward sloping dashed line is a counterfactual level of lending where sx = 0.
therefore, to increase the realised return to the marginal loan in all states. This
applies in the default region, which tends to increase the transfer to the taxpayer,
thereby making the overhang problem more acute. And it applies on the boundary
too, which tends to reinforce the FSE. Identifying these effects is not only useful
in helping to interpret the results of the sensitivity analysis in our calibration
exercise below, but it also tells us that other frictions that reduce the equilibrium
level of lending could have similar reinforcing effects.
4.3 The full model
We are now in a position to describe the general model. This model includes the
two extensions above: risk on new and legacy loans and the corporate income
tax. We also modify the capital requirement to takes into account that, in real
world regulation, the requirements apply to risk-weighted assets. Accordingly, we
27
rewrite the constraint as:
κ+ c ≥ γ(αx+ βλ),
where α and β are risk weights parameters.
So far, we have focused on a single bank, facing a given downward sloping
demand for loans. The loan demand was not affected by the capital requirement.
In practice, however, the loan demand for a bank is affected by the loan supply of
other banks and, therefore, by the capital requirements they face. A calibration
of the model that aims at informing the debate on the overall level of capital
requirements should take such bank interactions into account.
To do this, we first propose a Cournot competition extension of the baseline
model.21 We take aggregate demand for loans as given, and consider a given
number ν of identical banks that all face the same capital requirement γ. Up to a
normalisation that we will introduce later, the payoff function of the representat-
ive bank takes the form:
BX(x) = Bx((x+ x′)−η
),
where x′ captures the total lending by other banks (and is taken as given), and η
is a parameter that captures the elasticity of aggregate loan demand.
Our equilibrium concept is a symmetric Nash equilibrium. Assuming it is
unique, it corresponds to the fixed point (i.e. x = x′) that solves the representative
bank’s first order condition.22
5 Empirical relevance
Our benchmark calibration aims at capturing a plausible situation facing a major
international bank in 2017. Table 1 summarises this calibration.
21A Cournot approach is analytically convenient, but we also believe that it is particularlymeaningful if one thinks that banks first choose their level of capital (which, given the capitalrequirement creates a capacity constraint) and then compete in price (i.e., in interest rate) in themarket for loans. Schliephake and Kirstein (2013) have shown that this results in an elegantapplication of Kreps and Scheinkman (1983): the equilibrium outcome corresponds to that underCournot competition. Other papers using Cournot competition for banks include Corbae andD Erasmo (2017) and Jakucionyte and van Wijnbergen (2018).
22In our numerical explorations, we have not encountered multiple fixed points.
28
Tab
le1:
Ben
chm
ark
Cal
ibra
tion
Par
amet
erVal
ue
Defi
nit
ion
Cal
cula
tion
Sou
rce(
s)γ
0.1
3ca
pit
alre
quir
emen
tT
ier
1R
isk
Bas
edM
inim
um
Cap
ital
Req
uir
emen
tof
Glo
bal
lyS
yste
mic
ally
Impor
tan
t
Ban
ks.
BC
BS
(2017)-
Tab
leB
.4
α0.5
risk
wei
ght
onn
ewlo
ans
Ave
rage
risk
wei
ghts
.M
aria
thas
anan
dM
erro
uch
e(2
014)
β0.5
risk
wei
ght
onle
gacy
loan
s
xMM
1M
Mle
velof
len
din
gN
orm
alis
atio
n.
ρ0.0
12
inte
rest
rate
Ave
rage
1ye
arco
nst
ant
mat
uri
tyU
Str
easu
ry
yiel
d.
Fed
eral
Res
erve
Boa
rd-
Rel
ease
H.1
5
σA
0.0
41
stan
dar
ddev
iati
onoflog(A
)Tar
getp=
0.97:
ann
ual
freq
uen
cyof
ban
kin
g
cris
esin
OE
CD
cou
ntr
ies
1970-2
012.
Val
enci
aan
dLae
ven
(2012)
σB
0.0
41
stan
dar
ddev
iati
onoflog(B
)
µA
log(1
+r)-
0.5σ2 A
expec
tati
onoflog(A
)E[A
]=
1+ρ,ex
isti
ng
loan
sfa
irly
valu
ed.
µB
log(1
+r)-
0.5σ2 B
expec
tati
onoflog(B
)E[B
]=
1+ρ,im
pliesxM
M=
1.
λ4
boo
kva
lue
ofle
gacy
loan
sxM
Mn
orm
alis
ed,an
dxM
M/λ
=⇒
20%
of
loan
sm
atu
rin
gper
year
.
van
den
Heu
vel(2
009)
ν12
nu
mber
ofban
ks
Loa
nsp
read
over
dep
osit
rate
=2%
Ber
nan
ke
etal
.(1
999)
η0.2
inte
rest
elas
tici
tyof
dem
and
inte
rest
elas
tici
tyof
dem
and
onm
ortg
age
deb
t
esti
mat
edfr
omU
Klo
an-t
o-va
lue
not
ches
.
Bes
tet
al.(2
015)
τ0.2
4co
rpor
ate
tax
rate
OE
CD
aver
age
corp
orat
eta
xra
te2005/2017.
OE
CD
tax
dat
abas
e
29
Calibrating the capital requirement The capital requirement is not a straight-
forward object to calibrate. In the model, what matters is loss absorbing liabilities,
as a percentage of the bank’s assets. Even accounting for risk weights, this is not
necessarily the same object as the headline regulatory capital requirement that is
the focus of the policy debate.
Relevant considerations include: (i) Banks have hybrid liabilities that both
may or may not count towards the requirement and may or may not have impli-
cit guarantees attached to them; (ii) there are different requirements for different
types of capital; (iii) banks hold voluntary buffers above the requirements (for in-
stance, to prevent small shocks from leading to violations); and (iv) requirements
vary across jurisdictions, types of banks (for example, banks deemed to be glob-
ally systemic now face higher requirements), and with macroeconomic conditions
(this is the role of counter-cyclical capital buffers).
To circumvent the issue, we present our results for a wide range of values of
γ. That is, we display the x∗(γ) functions. Still, we need a reference value to
centre the calibration. For ease of interpretation, we consider a headline number
of 13% of risk weighted assets for the requirement. Under Basel III, this roughly
corresponds to the Tier 1 capital requirement (including systemic, conservation,
and pillar 2 buffers) that globally systemically important banks face (see BCBS
(2017) and EBA (2017) for recent assessments; Table A in BoE (2015) describes a
breakdown of different requirements). This is the number we use for our calibra-
tion.
Average risk weights are typically around 50% (see Mariathasan and Mer-
rouche (2014)), so this is the number we use for α and β.23 Again, in practice,
there is variation across banks, and over time.
Taxation and interest rates We calibrate τ to match the simple average stat-
utory corporate tax rate among OECD countries; this corresponds to 24% in
23Using the same value for α and β makes exploring a range of γ is equivalent to exploring arange for average risk weights.
30
2017. In our sensitivity analysis, we show the effect of cutting the tax rate from
35% to 21%, which are the percentage rates before and after the recent tax reform
in the US. We interpret the period in our model as one year. Hence, we calibrate
the interest rate ρ to match the average 1-year constant maturity US treasury
bond yield: 1.2% in 2017.
Normalisation, competition, and elasticity of demand. We select parameters
so that xMM = 1 in the benchmark calibration and any alternatives presented. To
this effect, we rescale the representative bank’s gross return function:
BX(x) = Bkx((x+ x′)−η
), (14)
where k = νη/(1− ην), and E[B] = 1 + ρ.
We calibrate η to match the interest elasticity of demand on residential mort-
gage debt estimated from UK loan-to-value notches (Best et al. (2015)). Last, we
choose ν to target the average spread on new loans in the model(E[B]X
∗
x∗− 1− ρ
),
which we calibrate at 2%, consistent with Bernanke et al. (1999). This gives
ν = 12.
We calibrate the book value of legacy assets (λ) to 4 such that, if the bank
chooses x = xMM , 20% of loans on the balance sheet were made in the current
period. This is in line with values in the literature (see, for example, van den
Heuvel (2009)). To abstract from bank closure (see Appendix A.1), we assume
that κ > γλ.
Risks and probability of default We model the joint distribution of f(A,B) as a
log-normal. We assume that legacy loans are held at fair value on the bank’s bal-
ance sheet; that is, E[A] = 1 + ρ. Finally, we assume that A and B have identical
standard deviations, which we calibrate by targeting the bank’s equilibrium de-
fault probability (1− p∗ in the model).
In the data, the appropriate calibration for 1 − p∗ is the probability that the
implicit subsidy is in the money and the bank benefits from a taxpayer transfer.
Determining the value of p from the price of bank securities is challenging, pre-
cisely due to the need to strip out the value of any expected transfer. Instead,
31
we use realised frequencies. Specifically, Valencia and Laeven (2012) find that
there have been 40 banking crises among the 34 OECD members over the period
1970-2012, which suggests a target value of p∗ = 0.97 or a 3% annual probability
of default (Martinez-Miera and Suarez (2014) use a similar value in their calibra-
tion).
Last, we set the correlation parameter between A and B equal to 0.5. This
choice is arbitrary, and we run sensitivity analysis over it below.
5.1 Results
Benchmark example Figure 7 displays the x∗(γ) curve (the left panel) and the
associated probability of survival p∗(γ) (right panel) for our benchmark calibration.
As in Figure 6, the downward sloping dashed curve filters out the direct effect of
s∗x on x∗(γ). The vertical dotted line indicates the reference value for γ.
As we can see, the representative bank is in an upward sloping region that
extends from a requirement of about 11% to 21%. At lower values of γ, the slope
is negative (and relatively steeper at very low values). From values 21% onward,
the curve is downward sloping again: the bank is in fact very safe, and what
dominates is the direct effect of taxation (as in Figure 6 in the subsection on
taxes).
Discussion and sensitivity analysis The slope is positive at our reference cap-
ital requirement, γ = 13%, but the response is economically small. For instance,
a capital requirement increase from 13% to 14%, would generate an increase in
lending of 0.02%. However, this is still very different from conventional wisdom
and the typical concern that such a policy change would provoke a cut in lend-
ing. And the reason for the absence of a cut is the forced safety effect. Here, it
is strong enough to more than offset the other forces in the model. The result
also contrasts with those of recent quantitative studies that suggest a strongly
negative lending response (e.g., Corbae and D Erasmo (2017) and Elenev et al.
(2017)).
Now consider an 8% capital requirement. The curve is downward sloping, and
32
Figure 7: Equilibrium Lending and Survival Probability Under the BenchmarkCalibration
(a) x∗(γ) (b) p∗(γ)
Notes: The two panels show equilibrium lending and survival probability for different levels of the capital requirementunder the benchmark calibration defined in Table 1. Panel a: equilibrium lending for the representative bank underthe alternative levels of γ; the downward sloping dashed line is the counterfactual level of lending where sx = 0; thevertical dashed line denotes the reference level of the capital requirement. Panel b: equilibrium survival probability forthe representative bank under the alternative levels of γ; the vertical dashed line denotes the reference level of the capitalrequirement.
steeper, which is more in line with conventional wisdom. Given that this percent-
age is the one mandated by Basel I, the regulation in place in most countries in
the 1990’s and most part of the 2000’s, this case constitutes a plausible situation
facing the banks before the global financial crisis (note also the corresponding,
higher probability of a bailout in the right panel).24 In this case, going from a
capital requirement of 7% to 8% causes a cut of lending of 0.2% and an increase
in lending spread of 5bps.
We believe that our benchmark numbers for the calibration are plausible.
However, they only constitute one example; we do not wish readers to take our
results as a prediction that raising capital requirements today would necessarily
cause most banks to increase lending. Nor would we want to claim, on the basis
of the paragraph above, that all banks would have shown steep negative lending
responses in the run-up to the global financial crisis.
We would rather argue that lending responses are likely to show a lot of vari-
24Given the many loopholes and the amount of regulatory arbitrage at the time, one could evenconsider an effective capital requirement below the headline 8%.
33
ation, in both the times series and in the cross section, and in both signs and
in magnitude. To illustrate this, Figure 8 shows how the lending response of the
representative bank in the benchmark example changes when we alter parameter
values one at a time. First, to generate a steeper positive response, one can, for
instance, make the legacy loans overvalued (see Panel a) or decrease the correl-
ation between A and B (Panel b). And vice versa: undervalued legacy loans and
higher correlation make the lending response more negative, and relatively steep
in some cases. Second, consistent with our analysis in Section 4.2, one can see in
Panel c that higher tax rates are generally associated with lower lending, but they
also affect the slope of the curve, as they tend to deepen the U-shape. Addition-
ally, higher interest rates tilt the curve clockwise, as tax deductibility becomes
more relevant.
5.2 Bank heterogeneity
So far, our numerical exercise has considered ν identical banks all facing an
identical capital requirement. In general, the lending of an individual bank is
more sensitive to an idiosyncratic change in its individual capital requirement
(see Figure 9 in Appendix C). This is simply because an individual bank faces a
much shallower residual demand curve for loans than the total banking system
(Kisin and Manela (2016) make a similar point).
It is also the case that heterogeneity among banks can substantially alter how
banks respond to a change in aggregate capital requirements. To illustrate this,
take one potential dimension in which banks could be heterogeneous: legacy
loan valuation. Imagine that, instead of all banks having fairly valued legacy
loans, as in the previous subsection, there are two equally sized groups of banks:
strong banks, whose assets are in fact better than their valuation (i.e. they are
undervalued) and weak banks that have overvalued legacy loans (or unrecognised
losses). Still, on average in the economy, legacy loans are fairly valued.
The two groups of banks will typically have different levels of lending. Broadly
speaking, strong banks will be safer and so should suffer from less of a guarantee
overhang, but they should also have less incentive to overlend. So either group
34
Figure 8: Sensitivity analysis on equilibrium lending
(a) Legacy Loan Valuation (b) Correlation
(c) Taxes (d) Interest Rates
Notes: The panels show equilibrium lending for the representative bank under the alternative levels of γ when the bench-mark calibration defined in Table 1 is modified in a single dimension. All panels: the blue line denotes benchmark calib-ration; the vertical dashed line denotes the reference level of the capital requirement. Panel a: the red dotted line is legacyloans undervalued by 1% (E(A) = 1.01(1 + ρ)), the red dashed line is legacy loans overvalued by 1% (E(A) = 0.99(1 + ρ)).Panel b: the red dotted line is when log(A) and log(B) have 0.8 correlation, red dashed line is when log(A) and log(B)have 0.2 correlation. Panel c: the red dotted line is τ = 0.21, the red dashed line is τ = 0.35. Panel d: the red dotted lineis ρ = 0.0, the red dashed line is ρ = 0.04.
35
could lend more than the other.
The banks will also respond differently to a change in their capital require-
ment. Put differently, each group has a different “U-shape” and, at a given γ,
these shapes have different slopes with potentially different signs. As we describe
below, competition further complicates the situation.
Such environment is highly non-linear. So aggregate lending, and the aggreg-
ate lending response to a change in the capital requirement, is affected by het-
erogeneity. In turns out that, compared to our representative bank benchmark
case, introducing heterogeneity can either magnify, mitigate, or even flip the sign
of the aggregate lending response.
Table 2 provides an example of how heterogeneity can play out. It shows the
change in lending response following a capital requirement increases of 1 and 2
percentage points respectively, with weak (strong) banks having legacy loans that
are 1% overvalued (undervalued). All other parameters are in line with Table 1,
and the initial capital requirement is 13%.
The responses for the representative bank are both small (they correspond to
our benchmark result, see Figure 7). It is immediately obvious that, with het-
erogeneity, the two groups of banks are each more sensitive to the requirement.
Our interpretation of what happens is the following. Weak banks are on the up-
ward sloping portion of their U-shape. Therefore, their individual response is to
increase lending. Strong banks’ legacy assets are less likely to generate negative
residual cash flows. Their FSE is weaker. For the sake of simplicity, assume that
their individual response is nil. Now we can consider the effect of competition.
The lending increase by weak banks decreases the residual demand facing strong
banks. As a result strong banks cut lending. But this increases the residual
demand facing weak banks, so they lend more (which also makes them safer and
feeds the FSE), and so on and so forth. In equilibrium, we end up with polarised
responses.
While this qualitative description is accurate for both columns in 2, between
the two cases the magnitudes and aggregate response differ markedly. When cap-
ital requirements are increased by 2 percentage points, this has a dramatic effect
on weak banks: they become much safer, and end up lending much more. Com-
36
Table 2: Heterogeneous lending responses to a capital requirement Increase
(1) (2)1% 2%
increase increaseRepresentative Bank Response (%) 0.021 0.034
Weak Banks: 1% Overvalued (%) 0.084 2.873Strong Banks: 1% Undervalued (%) -0.046 -0.642Aggregate Lending (%) -0.014 0.062
Notes: The table shows the response of lending to a 1 percentage point (column 1) and 2 percentage point (column 2)capital requirement increase starting at γ = 13%. The first row shows how the representative bank responds under thebenchmark calibration. The final three rows considers how the responses are altered by heterogeneity among banks.Specifically, we assume that half of banks have legacy loans that are 1% overvalued (E(A) = 0.99(1 + ρ)), and half havelegacy assets that are 1% undervalued (E(A) = 1.01(1 + ρ)). We denote these weak and strong banks respectively and thesecond and third columns present the lending response in each group. The final row presents the aggregated response oflending by both types of banks to the capital requirement increase. Note that at the initial value of γ = 13%, strong bankslend more hence the aggregate lending response places more weight on the reaction of strong banks. All other parametersare calibrated as in Table 1.
petition amplifies this further and we see a 2.9% increase in lending. While this
crowds out strong bank lending, the net effect is to boost aggregate lending by
0.06%, double the increase in the representative case. In contrast, a 1 percent-
age point increase gives a smaller boost to weak banks. Since strong banks are
initially on a downward sloping portion of their U-shape their response actually
dominates, and overall aggregate lending falls.
5.3 Link to the empirical literature
To sum up, the exercises above highlight that the determinants of bank beha-
viour under capital requirements are many and complex. There are circum-
stances where a bank’s lending will respond positively to a change in capital
requirements and, we have argued, these circumstances are neither extreme nor
implausible. Furthermore, a change in the bank’s circumstances could substan-
tially alter that response. One should not necessarily expect a stable relationship
between capital requirements and lending.
This may explain inconsistencies in the recent empirical literature. For ex-
ample, Gropp et al. (2018) and Bassett and Berrospide (2017) study stress test
37
induced increases in the capital requirements in Europe in 2011 and the United
States in 2013-2016, respectively. Both use types of difference-in-differences
estimators where size based stress test eligibility criteria determines treatment.
Despite the similar settings, the conclusions are very different: Gropp et al. (2018)
find a reduction in lending, and Bassett and Berrospide (2017) find that, if any-
thing, lending goes up. Yet both are consistent with our model: inspecting Figure
7, our calibration suggests that times where banks have low probability of failure
tend to be associated with flat or positive responses to capital requirements. In
contrast, the downward sloping region coincides with a higher default probability.
Given the stressed European banking sector facing the teeth of a sovereign debt
crisis in 201125 versus the recovering US in 2013-2016, our model perhaps helps
reconcile the two findings.
Jiménez et al. (2017) also find that wider economic conditions matter for the
lending response to capital requirement changes. Looking at reforms to dynamic
provisioning in Spain they find that aggregate credit supply contracts by less in
good times versus than bad. As with our model, this indicates that attempts to
extrapolate evidence from specific settings or time periods can be problematic.
Empirical evidence from pre-crisis sample periods generally points to a negat-
ive lending response (see for instance, Hancock and Wilcox (1994); Francis and
Osborne (2012); Aiyar et al. (2014a,b)). Our model predicts that a negative lend-
ing response is more likely at low levels of the capital requirement. It is, therefore,
conceivable that future empirical research, with sample periods under the stricter
requirements of the new Basel III regime, will have different findings.
Another strand of the empirical literature focuses on the response of different
types of lending to heterogeneous capital requirements or different risk weights.
For instance, Behn et al. (2016) provides evidence on whether capital require-
ments are binding and how certain types of borrowers would be affected. How-
ever, the response of specific asset holdings to their capital charge cannot be
extrapolated to the whole balance sheet. Looking at our model, imagine that the
bank could make two types of new loans, and that the risk weight was raised on
25De Jonghe et al. (2016) and Fraisse et al. (2017) also report that tighter capital requirementsreduced lending in European in samples during the crisis.
38
just one of them. The bank would do less of that type of lending relative to the
other type but if this makes the bank safer enough, the bank could still expand
its whole balance sheet, which would reflect a strongly positive FSE.
Finally, a related strand of the empirical literature focuses on the lending re-
sponse to shocks to the level of bank capital (Bernanke and Lown (1991); Ber-
rospide and Edge (2010) ) or to losses on the bank’s existing assets (Peek and
Rosengren (1997); Puri et al. (2011); Rice and Rose (2016)) . While these ques-
tions are related, they are different from asking how the capital requirement af-
fects lending. However, through the lens of our model we can say that an unre-
cognised loss to the bank’s existing assets (i.e., legacy loans are overvalued and
E(A) < 1) will typically reduce lending (see Figure 8a). A recognised loss, where
the book value of legacy loans is marked down to a fair value, is equivalent to
an equal reduction in κ and λ, holding E[A] fixed. However, as we will discuss
next, a change in κ, considered in isolation, typically has no effect on lending
in our model. Still, from an empirical standpoint, it is challenging to separate
recognised from unrecognised losses.
6 Residual cashflows
6.1 Residual cashflow heterogeneity
A key feature of our model is that, when valuing the marginal loan, the bank will
consider both its contribution to economic surplus and its effect on the implicit
subsidy. Put differently, the bank does not price assets independently, rather it
values them in relation to its whole balance sheet. In this subsection we discuss
the role of legacy loans, and how this pricing mechanism would play out under
alternative assumptions.
Legacy loans and the interpretation of λ We highlighted in Section 4 that
some residual cashflow heterogeneity is needed for the existence of a positive
lending response. In our model, this gives a key role to λ, which we interpret as
legacy loans. However, λ could equally be interpreted as a portfolio of securities,
39
or another business division. Note also, as we will discuss below, that our mech-
anism still applies if λ is exempt of capital charges (e.g., an OECD sovereign debt
exposure) or even if the exposure is not on the balance sheet (litigation risk for
instance).
Since banks are going concerns and have, at all times, a stock of existing
loans on the balance sheet, interpreting λ as legacy loans is a natural thing to
do. Nevertheless, we want to stress that it is not necessary that residual cashflow
heterogeneity comes from legacy loans.
Loan sales at fair price In our model, loans cannot be sold. What would happen
if they could be?
First, note that in the baseline model, the bank has no incentive to sell legacy
loans at a fair price (i.e., at a unit price E[A]). The reason is precisely that these
loans carry an implicit subsidy when they are held on the bank’s balance sheet.
By decreasing its exposure at the margin, the bank would simply reduce the value
of the subsidy. Selling all legacy loans is not in the bank’s interest either. It is
true that this would fully alleviate the guarantee overhang: without legacy loans
on the balance sheet, the bank would choose x∗ = xMM . But it would be better off
choosing x∗ = xMM while keeping the legacy loans.
This intuition is in line with the reluctance of a distressed bank’s shareholders
to clean up their balance sheet by selling risky assets at market prices (see, for
instance, Philippon and Schnabl (2013)).
When new loans are risky, selling legacy loans may become attractive for the
bank. However, since we assume legacy loans to be (among themselves) perfectly
correlated, the bank will either choose to keep them all, or to sell them all. Which
option is best depends on the overall risk structure. The logic of our analysis still
holds, though, and, at the margin, the effect of a change in the capital require-
ment on the bank’s choice will still be captured by the cross-partial derivative of
the subsidy.
Starting a new bank (i.e., selling the real option to lend) In our model, is-
suing new loans generally reduces the subsidy attached to legacy loans. If we
40
allowed the shareholders to start a second, separate bank in which they could
issue the new loans, they would do it. The reason is that, while economic surplus
is additive, the rents they get from government guarantees are not. In particu-
lar, the sum of the values of the two implicit subsidies exceeds the value of the
implicit subsidy on an integrated portfolio. This is basic option theory: the value
of the implicit subsidy is that of a put option (Merton (1977)), and the value of a
portfolio of options exceeds the value of an option on the portfolio.
6.2 Residual cashflow alignment
A conclusion from our calibration exercise is that whether a bank’s lending re-
sponse is positive or negative depends, on a complex way, on many variables.
Since these relationships are captured by the third cross-partial derivative of the
value of a put option, this is perhaps not too surprising. On the one hand, this
means that the model can help reconcile apparently conflicting findings. On the
other hand, this also means that the model does not provide a crisp set of em-
pirical predictions. Still, we can formulate a useful general prediction regarding
how government guarantees affect bank behaviour.
To do this, it is useful to think of the bank’s general asset pricing problem.
The bank’s relevant pricing kernel is binary: it values residual cashflows linearly
in the states it survives, and does not value them at all in the states it defaults.
As a result:
Financial institutions that benefit from government guarantees will over-value loans (and assets) that have negative residual cashflows in thestates of the world where the bank defaults; and will undervalue assetsthat have positive residual cashflows in those states.
This means that banks’ loan and asset portfolio decisions are distorted towards
aligning the sign of their residual cashflow. We refer to such behaviour as Resid-
ual Cashflow Alignment (RCA).
Menu of heterogeneous new loans What would RCA behaviour mean if the
bank could choose, with full flexibility, among a set of potential new loans? And
41
how would its choice be affected by the capital requirement?
Maintaining our assumption that legacy loans are not sold, what the bank
would do is to issue the loans whose residual cashflows in survival states are
at least γ% of face value, no matter their residual cashflows in default states.
As a result, depending on the joint distribution of payoffs, the marginal loan in
equilibrium (i.e., the one that the bank just did not make) may have positive or
negative NPV in equilibrium. Whether a change in capital requirement makes this
marginal loan more or less attractive would still crucially depend on the expected
residual cash flow along the default boundary.
The same logic would apply, in equilibrium, to a bank that has no legacy assets
on its balance sheet. Imagine such a bank faces a finite number of potential
loans. In equilibrium, the subset of loans the bank would choose to finance
would only include those with the residual cashflow structure described in the
previous paragraph. And, again, the marginal loan would have either positive or
negative NPV, and a change in capital requirement could make it more or less
attractive.
Harris et al. (2017) study a related portfolio allocation problem. In their model,
banks have different initial portfolios of loans, which are perfectly liquid, and
they compete to make new loans to heterogeneous borrowers. What the authores
find is that the general equilibrium allocation exhibits perfect residual cashflow
sign alignment. That is, assuming a bank holds a given loan in its portfolio
in equilibrium, then, it will have sold all loans (and not issued any loan) that
generate, in any state, residual cashflows of a different sign than that of this
loan.
In the author’s terminology, banks are seeking downside risk correlation (see
Landier et al. (2015) for a similar interpretation). We think that our terminology
is more accurate, as it is the alignment of the signs of the residual cashflow
that is relevant, not their correlation in default states. In fact, perfectly negative
correlation given default is compatible with perfect RCA.
Concrete examples of prediction Formulating our prediction in terms of RCA
is useful in several dimensions. First, we note that limited liability, in the pres-
42
ence of existing debt, also induces RCA behaviour. It is on this basis that we make
the link between risk-shifting, debt overhang, and guarantee overhang problems
in the introduction.
Second, building on four concrete cases that have been studied in the recent
literature, we can formulate scenarios in which we are relatively comfortable mak-
ing predictions on the sign of the lending response. (Note that the original papers
do not consider changes in capital requirements).
First, consider the case of New Century Finance Corporation (NC). Landier
et al. (2015) point out that the monetary tightening by the Fed in the spring of
2014 was a negative shock to NC’s sub-prime loan portfolio. They argue that, NC
reacted to this shock by resorting to deferred amortization loan contracts on a
massive scale, precisely because they would perform badly if house prices were
to fall. Imagine that a bank was in a similar situation; an increase in capital
requirement would mitigate the problem, and reduce excessively risky lending.
This is an intuitive case.
In general, however, banks have more heterogeneous assets on their balance
sheets than such a specialised mortgage originator. Our next two scenarios re-
flect this. Consider a scenario similar to that studied in Puri et al. (2011): a
German saving bank whose business is primarily to lend to domestic households
and SMEs. The bank has, however, a substantial exposure to US sub-prime
mortgages, which could generate sufficient losses to cause the bank to default
in some states. Assuming the marginal loan on the domestic portfolio would not
make dramatic losses in those states, our model predicts that the bank will cut
domestic lending as the outlook on US sub-prime worsens. This is consistent
with the findings of Puri et al. (2011). In addition, our model predicts that: (i) the
FSE is positive; and (ii) that it dominates (i.e., the lending response is positive).26
In our interpretation, the lending cut documented by Puri et al. (2011) reflects
RCA behaviour. In the literature, the scaling down of a bank that has faced an
adverse shock is sometimes interpreted as a sign of prudent behaviour, perhaps
reflecting an increase in risk aversion (e.g., DeYoung et al. (2015) and Elenev et al.
26This second point assumes that the subprime exposure is not too large.
43
(2017)).27 Our interpretation is fundamentally different: we predict that a bank
will pass on investment opportunities with misaligned residual cashflows. This
is because misaligned residual cashflows provide a form of (default) insurance,
whose benefits accrue to other stakeholders.28 Rather than prudence, this should
still be interpreted as a form of excessive risk taking. Furthermore, it is possible
to build examples where, in order to align residual cashflows, a bank passes
on what would appear in isolation to be a textbook risk-shifting opportunity.29
But again, what motivates such behaviour is not the aim of improving efficiency
or limiting risk exposure; the goal is simply to maximise expected payoff given
limited liability.
The third scenario is, in a sense, a variant of the second. Consider an Irish
bank that had been lending extensively to high LTV, urban mortgagors when the
Central Bank of Ireland imposed restrictions on the issuance of such loans. The
findings in Acharya et al. (2018) suggest that this policy would cause the bank
to aggressively expand its issuance of loans to safer borrowers, in rural counties.
The question then arises as to why the bank was passing on the second type
of loans in the first place (and why it is now willing to issue them at a lower
rate). One possibility is that the bank found these loans unattractive because
they would have produced positive expected residual cashflows in default states
in general, and on the boundary in particular. Then, the policy made the bank
safer, which made the bank internalise those positive cashflows. Our prediction is
27The argument is often based on Froot and Stein (1998), who have shown that, if raisingcapital is ex-post costly, banks will behave in a risk averse fashion (by building precautionarybuffers, for instance) in an effort to avoid those costs. Froot and Stein (1998) make however clearthat their analysis abstracts from the distortions created by government guarantees.
28Fecht et al. (2015) also find that deposit insurance hinders banks’ incentive to insure them-selves using the interbank market.
29Imagine that our bank faces a single new loan opportunity. This is a unit loan that repayseither 1.12 or 0.8, with equal probability. Imagine that the capital requirement is 10%. Eventhough it is negative NPV, a new bank would finance this loan, because investing 0.1 in capitalyields a payoff of 0.22 with 50% probability (and 0 otherwise). This would be a classic example ofrisk-shifting (Kareken and Wallace (1978)). But if the bank’s portfolio includes a legacy loan thatgenerates positive residual cashflows (of at least 0.02 in expectation) in the state where the newloan performs badly, the bank will refrain from making the loan. From the perspective of a newbank, the implicit subsidy makes the (intrinsically negative NPV) loan a positive NPV investment.But residual cashflow misalignment reduces the part of the payoff that is a rent (the marginalsubsidy).
44
that such loans would also have become more attractive after an overall increase
in capital requirements.
Finally, consider a market maker in a particular security business. These
securities are relatively safe (e.g., US treasuries) and not closely related to the rest
of the bank’s balance sheet. RCA behaviour implies that the bank may pass on
positive NPV trades, which is in line with the interpretation in Duffie et al. (2018).
Assuming the bank is not in distress to start with, an increase in the capital
requirement is likely to make the bank more willing to deal in those assets, and
should narrow the bid ask spread.
We find these examples useful for illustrating the potentially wide range of
applications for our results. Besides the concrete empirical predictions on the
sign of the lending response, a key takeaway is that, while RCA reflects excessive
risk-taking, the induced behaviour does not necessarily take the form of textbook
risk-shifting behaviour.30 This may help explain the lack of evidence that banks
start lending more aggressively as their balance sheets deteriorate.
6.3 Multiple requirements
Imagine that the bank faces two capital requirements γ1 and γ0, such that:
κ+ c ≥ γ1x+ γ0λ.
We can now redefine the residual cashflows from the marginal loan as Z = BXx−(1 − γ1). The expression for the marginal subsidy is then unchanged except that
the default boundary is altered.31 As with a single capital requirement, the sign
of the lending response depends on the derivative of the marginal subsidy with
30This observation is related to Gollier and Pratt (1996). They show that, in the presence ofbackground risk, a concave but piece-wise linear objective function can induce behaviour that ishard to interpret in terms of risk-aversion. In our context, limited liability makes the objectivefunction convex and piece-wise linear. And our example in the footnote above is indeed the mirrorof that proposed by Gollier and Pratt (1996).
31Specifically, we have a0(B) = 1λ ((1− γ1)x+ (1− γ0)λ−BX) and b0(A) =
1X ((1− γ1)x+ (1− γ0)λ−Aλ) .
45
respect to the requirement in question. Specifically, we have that for a change γ1:
s∗xγ1 = −(1− p∗) + p∗γ1z∗∆0.
The first term can be interpreted as a composition effect; the second term is a
form of FSE. Starting from γ1 = γ0 = γ, the only difference in this expression
from equation (13) is that p∗γ > p∗γ1 > 0. Raising the capital requirement on new
loans generates a smaller shift in the default boundary than an increase in the
requirement on all loans. The FSE is weakened but still present. Furthermore,
the expression is unchanged if γ0 = 0; in particular, the FSE can still be positive
and dominate the composition effect even if no capital is held against legacy loans
at all.32
Now consider a change in γ0:
s∗xγ0 = p∗γ0z∗∆0.
There is no composition effect: changing γ0 has no impact on the composition
of liabilities used to fund new loans. The sign of the lending response is de-
termined solely by the (modified) FSE; whether or not the bank increases lending
depends on the sign of the residual cashflows, in expectation, along the default
boundary.
Now consider what happens if γ1 = 1 such that new loans are fully financed
by capital. This brings us very close to the classic debt overhang model of Myers
(1977). The bank has existing assets that are partly debt financed (at a fixed in-
terest rate, in our model, due to the guarantee), but new investments are wholly
financed by capital. Now the (modified) FSE and the overhang problem go hand
in hand. As residual cashflows are always positive, Z = BXx, the bank always
underlends (s∗x = −(1− p∗)z∗∆ < 0), and lending is always increasing in γ0. Further-
more, the tighter requirement increases lending precisely because it makes the
overhang problem less severe.33
32Compared to γ0 = γ1 > 0, setting γ0 = 0, influences the equilibrium values of p∗, p∗γ1 and z∗∆0. It
will unambiguously make the bank riskier, strengthening the composition effect but not the factthat the default boundary shifts outward could increase both p∗γ1 and z∗∆0
, strengthening the FSE.33This result stands in sharp contrast with the conclusion of Admati et al. (2018) that a firm
46
This extreme case serves to highlight some of the original features of our
model. To generate the competing forces that are present in our model, one
must account for the fact that capital requirements inherently influence both the
residual cash flows on the marginal loan and the default boundary. This is why
we can have, emerging from the same moral hazard problem, ambiguous predic-
tions over the sign of the lending response, the sign of the FSE, and a disjunction
between the two. This example also illustrates that even if the marginal new loan
has an exactly identical payoff distribution as the average loan on the bank’s
balance sheet, one can still obtain residual cashflow heterogeneity if the capital
requirements (or the risk weights) on different assets are heterogeneous.
7 Conclusion
We see our contribution to the policy debate as follows: that capital is costly for
banks does not imply a negative lending response. Indeed, the Forced Safety Ef-
fect can counteract the liability composition effect, which overturns conventional
wisdom.The policy debate concerns both the long term effect of capital requirements
(what is their socially optimal level?) and their effect in the shorter term (how
should they be adjusted to the state of the economy?).
To maintain as much tractability as possible, we study a static model. En-
dowing the banks with legacy loans allows us to give banks a going-concern di-
mension. Crucially, it enables us to show that the way in which banks react to
capital requirements is history dependent: existing assets on their balance sheet
matters. Hence, our approach directly speaks to changes in capital requirements
(which can be interpreted as time varying adjustments, such as Basel III’s coun-
that faces a debt overhang problem and is forced to deleverage is more likely to sell assets thanto raise capital. A key difference is that they consider zero NPV trades in an environment thatis, essentially, scale invariant. Our approach allows us to study the effect of forced deleveragingstarting from a level that is privately optimal for the firm, given the initial leverage restriction.When there is a guarantee overhang, the marginal trade (i.e., the marginal loan) has a strictlypositive NPV. So, at the margin, an increase in capital requirement is unlikely to make a zero NPVloan attractive. However, given that the (modified) FSE is positive, it would make the (initially)marginal loan more attractive.
47
tercyclical buffers, or as gradual increases towards higher levels). Insights from
our paper could, for instance, inform a regulator who wants to time increases in
capital requirements in a way that minimises the impact on economic activity.
Our approach does not allow us to directly address the questions linked to the
overall long term ideal level of capital requirements. Given that lending responses
exhibit strong non-linearities, calibrating our model to average conditions and
interpreting this as a proxy for a steady state would not be appropriate.34 In
fact, in light of our results, it is hard to think of a relevant concept of steady
state. The fact that banks are going concerns, at any given time (excluding when
the bank starts), they will always have legacy loans. How the structure of their
residual cashflows relates to that of the marginal loan is a complicated question.
All objects reflect endogenous decisions, and the interaction also depends on the
regulatory regime.
From that point of view, solving a dynamic version of our model with het-
erogeneous banks could help draw new insights on the long-term and dynamic
effects of capital requirements. However, keeping such a model tractable will be
challenging. We leave this for future research.
ReferencesAcharya, V. V., Bergant, K., Crosignani, M., Eisert, T., McCann, F., 2018. The anatomy of
the transmission of macroprudential policies: Evidence from ireland, mimeo.
Admati, A. R., DeMarzo, P. M., Hellwig, M. F., Pfleiderer, P., 2013. Fallacies, irrelevantfacts, and myths in the discussion of capital regulation: Why bank equity is not sociallyexpensive. Stanford Graduate School of Business, Working Paper No. 2065.
Admati, A. R., DeMarzo, P. M., Hellwig, M. F., Pfleiderer, P. C., Feb. 2018. The leverageratchet effect. Journal of Finance 73 (1), 145–198.
Aiyar, S., Calomiris, C. W., Hooley, J., Korniyenko, Y., Wieladek, T., 2014a. The interna-
34De Nicolò et al. (2014) simulate a dynamic model of a bank in partial equilibrium. To proxyfor steady state, they look at unconditional moments. They find that lending is, on average, lowerat a 12% capital requirement than at 4%. They also find that lending is lower for an unregulatedbank. But this result is hard to interpret, as the bank operates, on average, with a negative bookvalue of equity.
48
tional transmission of bank capital requirements: Evidence from the u.k. Journal ofFinancial Economics 113 (3), 368–382.
Aiyar, S., Calomiris, C. W., Wieladek, T., 2014b. Does macro-pru leak? evidence from auk policy experiment. Journal of Money, Credit and Banking 46 (1), 181–214.
Allen, F., Gale, D., 2000. Bubbles and crises. Economic Journal 110 (460), 236–55.
Bahaj, S., Bridges, J., Malherbe, F., O’Neill, C., Apr. 2016. What determines how banksrespond to changes in capital requirements? Bank of England Staff Working Paper No.593.
Bassett, W. F., Berrospide, J. M., Sep. 2017. The impact of stress tests on bank lending,mimeo.
BCBS, Dec. 2017. Basel iii monitoring report: Results of the cumulative quantitativeimpact study. Basel Committee on Banking Supervision.
Begenau, J., Mar. 2018. Capital requirements, risk choice, and liquidity provision in abusiness cycle model, mimeo, Stanford University.
Behn, M., Haselmann, R., Wachtel, P., 04 2016. Procyclical capital regulation and lend-ing. Journal of Finance 71 (2), 919–956.
Bernanke, B. S., Gertler, M., Gilchrist, S., 1999. The financial accelerator in a quant-itative business cycle framework. In: Taylor, J. B., Woodford, M. (Eds.), Handbook ofMacroeconomics. Vol. 1 of Handbook of Macroeconomics. Elsevier, Ch. 21, pp. 1341–1393.
Bernanke, B. S., Lown, C. S., 1991. The credit crunch. Brookings Papers on EconomicActivity 22 (2), 205–248.
Berrospide, J. M., Edge, R. M., Dec. 2010. The effects of bank capital on lending: whatdo we know, and what does it mean? International Journal of Central Banking 6 (4),5–54.
Best, M. C., Cloyne, J., Ilzetzki, E., Kleven, H. J., Aug. 2015. Interest rates, debt andintertemporal allocation: evidence from notched mortgage contracts in the united king-dom. Bank of England Staff Working Paper No. 543.
BoE, Dec. 2015. Supplement to the december 2015 financial stability report: The frame-work of capital requirements for uk banks. Bank of England.
Brooke, M., Bush, O., Edwards, R., Ellis, J., Francis, B., Harimohan, R., Neiss, K.,Siegert, C., Dec. 2015. Measuring the macroeconomic costs and benefits of higher ukbank capital requirements. Financial Stability Paper 35, Bank of England.
49
Corbae, D., D Erasmo, P., 2017. Capital requirements in a quantitative model of bankingindustry dynamics. mimeo, University of Wisconsin.
De Jonghe, O., Dewachter, H., Ongena, S., Oct. 2016. Bank capital (requirements) andcredit supply: Evidence from pillar 2 decisions. Research Working Paper 303, NationalBank of Belgium.
De Nicolò, G., Gamba, A., Lucchetta, M., 2014. Microprudential regulation in a dynamicmodel of banking. Review of Financial Studies 27 (7), 2097–2138.
DeYoung, R., Gron, A., Torna, G., Winton, A., Sep. 2015. Risk overhang and loan port-folio decisions: Small business loan supply before and during the financial crisis. TheJournal of Finance 70 (6), 2451–2488.
Duffie, D., Andersen, L., Song, Y., Mar. 2018. Funding value adjustments. Journal ofFinance Forthcoming.
EBA, Dec. 2017. Ad hoc cumulative impact assessment of the basel reform package.European Banking Authority.
Elenev, V., Landvoigt, T., van Nieuwerburgh, S., Sep. 2017. A macroeconomic model withfinancially constrained producers and intermediaries. CEPR Discussion Papers 12282,C.E.P.R. Discussion Papers.
Estrella, A., June 2004. The cyclical behavior of optimal bank capital. Journal of Banking& Finance 28 (6), 1469–1498.
Fecht, F., Inderst, R., Pfeil, S., 2015. A theory of the boundaries of banks with implica-tions for financial integration and regulation. IMFS Working Paper Series 87, GoetheUniversity Frankfurt, Institute for Monetary and Financial Stability (IMFS).
Fraisse, H., Lé, M., Thesmar, D., Jun. 2017. The real effects of bank capital requirements.ESRB Working Paper Series 47, European Systemic Risk Board.
Francis, W. B., Osborne, M., 2012. Capital requirements and bank behavior in the uk:Are there lessons for international capital standards? Journal of Banking & Finance36 (3), 803–816.
Froot, K. A., Stein, J. C., 1998. Risk management, capital budgeting, and capital struc-ture policy for financial institutions: An integrated approach. Journal of Financial Eco-nomics 47 (1), 55 – 82.
Gollier, C., Pratt, J. W., 1996. Risk vulnerability and the tempering effect of backgroundrisk. Econometrica 64 (5), 1109–23.
Gropp, R., Mosk, T., Ongena, S., Wix, C., 2018. Banks response to higher capital require-ments: Evidence from a quasi-natural experiment. The Review of Financial Studies.
50
Hancock, D., Wilcox, J. A., 1994. Bank capital and the credit crunch: The roles of risk-weighted and unweighted capital regulations. Real Estate Economics 22 (1), 59–94.
Hanson, S. G., Kashyap, A. K., Stein, J. C., 2011. A macroprudential approach to finan-cial regulation. The Journal of Economic Perspectives 25 (1), 3–28.
Harris, M., Opp, C. C., Opp, M. M., Aug. 2017. Bank capital and the composition ofcredit, mimeo.
IIF, Sep. 2011. The cumulative impact on the global economy of changes in the financialregulatory framework. Institute of International Finance.
Jakucionyte, E., van Wijnbergen, S., Jan. 2018. Unclogging the credit channel: Onthe macroeconomics of banking frictions. Tinbergen Institute Discussion Papers 18-006/VI, Tinbergen Institute.
Jensen, M. C., Meckling, W. H., 1976. Theory of the firm: Managerial behavior, agencycosts and ownership structure. The Journal of Financial Economics 3, 305–360.
Jiménez, G., Ongena, S., Peydró, J.-L., Saurina Salas, J., Dec. 2017. Macroprudentialpolicy, countercyclical bank capital buffers and credit supply: Evidence from the span-ish dynamic provisioning experiments. Journal of Political Economy, 125 (6), 2126–2177.
Kareken, J., Wallace, N., 1978. Deposit insurance and bank regulation: A partial-equilibrium exposition. Journal of Business, 413–438.
Kisin, R., Manela, A., 2016. The shadow cost of bank capital requirements. Review ofFinancial Studies 29 (7), 1780–1820.
Kreps, D. M., Scheinkman, J. A., 1983. Quantity precommitment and bertrand competi-tion yield cournot outcomes. The Bell Journal of Economics 14 (2), 326–337.
Krugman, P., 1999. What happened to asia. In: Sato R., Ramachandran R.V., M. K. (Ed.),Global Competition and Integration. Research Monographs in Japan-U.S. Business &Economics. Vol. 4. Springer, Boston, MA.
Landier, A., Sraer, D., Thesmar, D., June 2015. The risk-shifting hypothesis: Evidencefrom subprime originations, mimeo.
Malherbe, F., 2015. Optimal capital requirements over the business and financial cycles.CEPR Discussion Paper No. DP10387.
Mariathasan, M., Merrouche, O., 2014. The manipulation of basel risk-weights. Journalof Financial Intermediation 23 (3), 300–321.
51
Martinez-Miera, D., Suarez, J., Sep. 2014. Banks endogenous systemic risk taking.mimeo, CEMFI.
McKinnon, R., Pill, H., 1997. Credible economic liberalizations and overborrowing. Amer-ican Economic Review 87 (2), 189–93.
Merton, R., 1977. An analytic derivation of the cost of deposit insurance and loan guar-antees an application of modern option pricing theory. Journal of Banking and Finance1, 3–11.
Modigliani, F., Miller, M. H., 1958. The cost of capital, corporation finance and the theoryof investment. American Economic Review 48 (3), 261–297.
Myers, S. C., 1977. Determinants of corporate borrowing. Journal of Financial Economics5 (2), 147–175.
Peek, J., Rosengren, E. S., 1997. The international transmission of financial shocks: Thecase of japan. American Economic Review 87 (4), 495–505.
Philippon, T., Schnabl, P., 2013. Efficient recapitalization. The Journal of Finance 68 (1),1–42.
Puri, M., Rocholl, J., Steffen, S., June 2011. Global retail lending in the aftermath ofthe US financial crisis: Distinguishing between supply and demand effects. Journal ofFinancial Economics 100 (3), 556–578.
Repullo, R., Suarez, J., 2013. The procyclical effects of bank capital regulation. Review ofFinancial Studies 26 (2), 452–490.
Rice, T., Rose, J., 2016. When good investments go bad: The contraction in communitybank lending after the 2008 gse takeover. Journal of Financial Intermediation 27 (C),68–88.
Rochet, J., 1992. Capital requirements and the behaviour of commercial banks.European Economic Review 36 (5), 1137–1170.
Schliephake, E., Kirstein, R., 06 2013. Strategic effects of regulatory capital requirementsin imperfect banking competition. Journal of Money, Credit and Banking 45 (4), 675–700.
Suarez, J., 2010. Discussion of: Procyclicality of capital requirements in a general equi-librium model of liquidity dependence. International Journal of Central Banking 6 (4),175–186.
Thakor, A. V., 1996. Capital requirements, monetary policy, and aggregate bank lending:theory and empirical evidence. The Journal of Finance 51 (1), 279–324.
52
Valencia, F., Laeven, L., Jun. 2012. Systemic banking crises database; an update. IMFWorking Papers 12/163, International Monetary Fund.
van den Heuvel, S., 2009. The bank capital channel of monetary policy. Discussion paper,Federal Reserve Board.
A Additional Theoretical Results
A.1 Participation constraint and bank closure
In our model, the shadow value of initial capital is equal to the price of new
capital. It is therefore irrelevant how much capital comes from new shareholders
versus how much was already on the bank’s books. This is why the value of κ
does not affect the value of x∗.
Still, initial shareholders have the option to close the bank at date 1 and walk
away with zero. This means that κ alters their participation constraint. To see
this, consider again the baseline model, the participation constraint is:
w∗ =
∫ aH
aL
(X∗ − x∗) f(A)dA+
∫ aH
aL
(Aλ− λ)) f(A)dA+ s∗ + κ ≥ 0. (15)
Under our assumption that E[A] = 1, the second term disappears, and the con-
straint is always satisfied (all other terms are positive). However, if E[A] < 1, κ
becomes relevant, and the participation constraint can be violated. First, con-
sider κ ≥ γλ. Here, closing the bank is never the best option for shareholders.
This is because, under limited liability, even operating at x = 0 gives the bank’s
shareholders a positive payoff in expectation. However, if κ < γλ, shareholders
must first raise new capital if the bank is to operate. When the option value of
operating the bank is low (i.e., when E[A] is low and new loans do not gener-
ate much surplus), operating may not be worth the cost of recapitalisation. The
53
participation constraint is then violated.35
Now, the key point we want to make is that when E[A] < 1, increasing γ may
make the bank close at date 1. Formally:
Proposition 5. If κ + (X∗ − x∗) < (1− E[A])λ, there exists a γ < 1 such that for allγ ≥ γ, the bank closes at date 1.
To understand Proposition 5, first note that a high γ makes it more likely that
κ < γλ. Second, γ reduces the subsidy (see Equation 7), so that s∗ = 0 for some
sufficiently large γ. Given s∗ = 0, if initial equity plus the surplus on new loans is
insufficient to cover the expected losses on legacy assets, then the bank will shut
down.
Since a large γ is not viable for distressed banks, the logic behind Proposition
1 does not necessarily apply here. If x∗(γ) is only upward sloping when γ > γ, the
bank would always choose to close rather than increase lending in response to a
requirement increase.
A.2 The model with taxes
Consider the baseline version of the model. Adding the tax and assuming that
households have an opportunity cost of funds 1+ρ > 1 means the bank’s objective
function now reads:
w = X − (1 + ρ)x︸ ︷︷ ︸economic surplus
+
∫ a0
aL
((1− γ) (1 + ρ) (x+ λ)−X − Aλ) f(A)dA︸ ︷︷ ︸≡s(x,γ), i.e. the implicit subsidy
− τ∫ aH
a1
(X + Aλ− (1 + (1− γ)ρ)(x+ λ)) f(A)dA︸ ︷︷ ︸≡t(x,γ), i.e. the expected tax bill
+κ,
35To see this, rewrite the participation constraint as:∫ aH
a0
(X∗ +Aλ− (1− γ)(x∗ + λ)) f(A)dA︸ ︷︷ ︸option value of operating
≥ (x∗ + λ)γ − κ︸ ︷︷ ︸new capital needed
.
54
where a1 = 1λ
((1 + (1− γ)ρ)(x+ λ)−X) denotes the threshold such that the bank
has positive taxable income if A > a1. The expected tax bill is then given by t(x, γ).
All other terms are in line with Section 3, except that they now account for ρ.
The tax adds a new wedge in the first order condition:
X∗x − (1 + ρ) + s∗x − t∗x = 0, (16)
where tx = qτ (Xx − (1 + (1− γ)ρ)) with q =∫ aHa1
f(A)dA is the probability that the
bank pays tax. Intuitively, t∗x > 0. Ceteris paribus, the tax reduces equilibrium
lending.
Now, the question we are interested in is how the tax, and the tax shield, affect
the shape of the lending response. The reasoning follows the same logic as before:
what happens depends on the cross-partial derivatives. But now, we have:
w∗xγ = s∗xγ − t∗xγ.
The introduction of a tax has two effects on w∗xγ:36 a direct effect that mater-
ialises through the appearance of a term t∗xγ, and an indirect effect that works
through a change in the value of s∗xγ.
To understand how these two different effects play out, first inspect Figure
6. The solid red line depicts x∗(γ) without the tax (τ = 0). This is the U-shape
relationship of the baseline model. The blue dashed curve is the case with the
tax. The net effect of taxes on the lending response is ambiguous: the U-shape
relationship is still present, but it is (i) tilted clockwise and (ii) deeper than in the
absence of a tax.
The tilt The tilts emerges because t∗xγ > 0.37 Intuitively, the wedge driven by tax-
ation is increasing in the capital requirement due to the deductibility of interest
on deposits. As a benchmark, note that setting sx = 0 (i.e., using X∗x−(1 + ρ)−t∗x = 0
as a counterfactual equilibrium condition) would yield the downward slopping
36The tax also affects w∗xx, and therefore, the magnitude of the lending response. As before, itssign is solely determined by w∗xγ.
37t∗xγ = q∗τρ+ q∗γτ (X∗x − (1 + ρ(1− γ)) > 0
55
dotted curve . It then follows that, as γ gets sufficiently high and the bank be-
comes safe, the level of lending converges to this line. Comparing this line to the
xMM level of lending makes the tilt more evident.
The deeper U-shape The deeper U-shape is due to the indirect effect of the tax
on s∗xγ. As in the baseline model, we have:
s∗xγ = −(1− p∗) + p∗γZ∗,
but the presence of the tax reduces x∗, which affects its equilibrium value. In
the baseline model, we argued that the implicit subsidy works as a tax at the
margin. The key implication was that it reduced equilibrium lending and made
the marginal loan a positive contributor to the bank’s residual cash flow. Now we
have introduced an actual tax, which reduces lending further and reinforces this
mechanism.
This has three effects. First, less (positive NPV) lending reduces profitability,
which increases the probability of default (1−p∗). Second, it affects p∗γ. Third, and
crucially, it raises the equilibrium marginal return to lending X∗x, which therefore
raises Z∗ and makes the forced safety effect stronger. The net effect is to deepen
the U-shape. At low levels of γ, the presence of the tax makes the bank respond
more negatively to an increase in γ, but at higher levels, it makes it respond more
positively. As the bank becomes fully safe, s∗xγ = 0, and only the direct effect is at
work. The response becomes negative again.
The tax and tax shield It is important to differentiate between taxes and tax
shields. Absent the deductibility of interest payments, txγ = 0, and the tax only
matters for the lending response through its impact on s∗xγ. Now, introducing a
tax shield on deposits means both that t∗xγ > 0 and that the marginal tax rate, t∗x,
is lower, which in turn potentially reduces the impact of the tax on s∗xγ.
Adding a second source of risk Combining the corporate income tax (and the
tax shield) with two sources of risk creates additional effects. Given that the
56
expression for t(x, γ) is very similar to that for s(x, γ), the derivations are very
similar, and it should not be too surprising that counter-intuitive cases can arise
for some sets of parameter values. In particular, it is possible that t∗xγ < 0 (which
would then contribute to a positive lending response) and t∗x < 0.38 However, these
are specific cases that are unlikely to be relevant in practice: in our calibration
exercise of Section 5, where there are two sources of risk, the effect of the tax
on x∗(γ) can essentially be understood and the basis of the tilt and the deeper
U-shape that we explain above.
B Proofs
Proposition. 1. For all γ ∈ (0, 1) and an associated x∗(γ) , if p (x∗(γ), γ) < 1, thereexists γ′ > γ such that x∗(γ′) > x∗(γ).
Proof. First, set γ′ = 1, and note that x∗(γ′)
= xMM . Hence, it suffices to establish
that ∀γ ∈ (0, 1), x∗(γ) < xMM . If w(x) is continuously concave, Problem 4 is convex
and the proof is straightforward. The first order condition is:
∂a0
∂x(X∗ + a∗0λ− (1− γ) (x∗ + λ))︸ ︷︷ ︸
=0
+
∫ aH
a∗0
[X∗x − (1− γ)] f(A)dA− γ = 0.
Rewriting this as X∗x = (1 − γ) + γ/p∗ makes clear that ∀γ ∈ (0, 1), X∗x > 1 if p∗ < 1,
which establishes the result.
However, w(x) may not be well behaved. For a more general proof, we exploit
the properties of the objective function on either side of a threshold x̂ such that
Xx(x̂) = 1− γ.
First consider x < x̂. Write the objective function as in Lemma 1 below, and
note that X − x is smooth and has a unique maximum at xMM . Note that for all
38A negative marginal tax requires that the marginal loan transfers cashflows from states wherethe rest of the bank’s assets are generating a tax profit to states where there is a tax loss. Hence,the marginal loan then reduces the expected tax bill and, absent other frictions, the tax can leadto the bank making negative NPV loans. In turn, a change in the capital requirement shifts theboundary between a taxable losses and profits. If the marginal loan generates a large tax loss onthis boundary, then it is possible that t∗xγ < 0.
57
x < x̂, and γ ∈ (0, 1), both the upper limit of integration a0(x), and the integrand,
are strictly decreasing functions of x. It follows that a point in the set [xMM , x̂[
cannot be a maximum of w(x).
Second, consider x ≥ x̂. Write the objective as stated in Problem (4). Now,
a0(x) is the lower limit of integration and is increasing in x. The integrand is also
decreasing in x, and so is the third term. Hence, w(x) is decreasing over the set
[x̂,∞. Hence, for all γ ∈ (0, 1), if p∗ < 1 then x∗(γ) < xMM .
Lemma. 1. The objective function can be rewritten:
w(x) = X − x︸ ︷︷ ︸economic surplus
+
∫ a0
aL
((1− γ) (x+ λ)−X − Aλ) f(A)dA︸ ︷︷ ︸≡s(x,γ), i.e. the implicit subsidy
+κ (17)
Proof. Use the assumption that E[A] = 1 and rearrange the objective function in
equation (4).
Proposition. 2. (The sign of the lending response)
dx∗
dγS 0⇔ s∗xγ S 0.
Proof. This result follows directly from the implicit function theorem applied to
the first order condition.
Proposition. 3 (i) z∗∆0can be positive; (ii) this implies a positive forced safety effect;
(iii) and can lead to a positive lending response: s∗xγ > 0.
Proof. We provide examples of positive lending responses in Section 5. Upon
request, we can provide examples with τ = 0 (see also Appendix C). Now, having
proved (iii) by example, (ii) must be true, and, therefore, (i) ,as well, since the
composition effect is always negative.
Proposition. 4 For all γ such that s∗x(γ) < 0, then there exists γ′ ≥ γ such that
x∗(γ′) > x∗(γ).
Proof. s∗x(γ) < 0⇒ x∗(γ) < xMM . Then, note that x∗(γ′= 1)
= xMM .
58
Figure 9: The lending response for an individual bank versus the banking system
Notes: We use the benchmark calibration as in Table 1. The blue line shows results for x∗(γ) assuming all banks face thesame capital requirement under the benchmark calibration. The red line shows the optimal lending decision for a singlebank facing a change in its own capital requirement, and we modify equation (14) such that x
′(lending at other banks) is
held fixed at x∗(γ = 0.13).
Proposition. 5. If κ + (X∗ − x∗) < (1− E[A])λ, there exists a γ < 1 such that for allγ ≥ γ the bank closes at date 1.
Proof. Since X′(0) = ∞, if the bank decides to operate, it must be the case
that x∗ > 0. Given that in equilibrium the marginal loan satisfies X∗x = (1 −γ) + γ
p∗> 1, all infra-marginal loans have positive net present value. Hence,∫ aH
aL(X∗ − x∗) f(A)dA > 0. Since new loans generate some surplus, there exists
a γ < 1 such that p∗ = 1. But then, s∗ = 0, and the participation constraint
(15) simplifies to (X∗ − x∗) + (E[A]− 1)λ + κ ≥ 0. The condition in the Proposition
ensures that it is violated.
C Additional figures
In our calibration exercise in Section 5, we run comparative statics over a sym-
metric equilibrium for ν identical banks facing the same capital requirement. As
an alternative, Figure 9 presents, in red, the optimal level of lending for an in-
dividual bank when its requirement is changed, assuming all other banks lend
59
Figure 10: Example of a positive lending response when the bank overlends
Notes: This figure plots x∗(γ) for an example in which the bank always overlends relative to xMM but the equilibriumlevel of lending is increasing in γ for some range. Some pertinent details of the calibration are provided in the main text.In addition, we calibrate the parameters as follows: η = 0.2; ν = 1; λ = 1; τ = 0; ρ = 0; µA = 1; σ2
A = 0.063; σ2B = 0.003;
and µB is set such that xMM = 1.
at a rate given by the initial equilibrium when γ = 0.13. The blue line shows the
symmetric equilibrium for comparison. As can be seen, the red curve is much
steeper.
In Section 4.1, we claimed it was possible to generate examples where the
lending response is positive and the bank overlends relative to xMM . Figure 10
presents one such example. The crucial deviation from our benchmark calibra-
tion is that we assume that when A realises below its median value, there is a 1%
chance that B = 0. Otherwise, the two random variables are distributed jointly log
normal. We adjust the means such that expected values are not affected (specific-
ally, xMM = 1 in this example) and we set the correlation between the two to -0.9.
The net effect of this change is to push some probability mass into very bottom
corner of the default region in Figure 5. So A and B have high tail dependence
but are negatively related for most of the A×B domain. The high tail dependence
means that z∗∆ < 0: there is a set of default states where the marginal new loan
performs very badly. But these states are typically not on the default boundary,
so it can be true that z∗∆0> 0.
60